Daily Papers

To standardize interactions between LLM-based agents and their environments, the Model Context Protocol (MCP) was proposed and has since been widely adopted. However, integrating external tools expands the attack surface, exposing agents to tool poisoning attacks. In such attacks, malicious instructions embedded in tool metadata are injected into the agent context during MCP registration phase, thereby manipulating agent behavior. Prior work primarily focuses on explicit tool poisoning or relied on manually crafted poisoned tools. In contrast, we focus on a particularly stealthy variant: implicit tool poisoning, where the poisoned tool itself remains uninvoked. Instead, the instructions embedded in the tool metadata induce the agent to invoke a legitimate but high-privilege tool to perform malicious operations. We propose MCP-ITP, the first automated and adaptive framework for implicit tool poisoning within the MCP ecosystem. MCP-ITP formulates poisoned tool generation as a black-box optimization problem and employs an iterative optimization strategy that leverages feedback from both an evaluation LLM and a detection LLM to maximize Attack Success Rate (ASR) while evading current detection mechanisms. Experimental results on the MCPTox dataset across 12 LLM agents demonstrate that MCP-ITP consistently outperforms the manually crafted baseline, achieving up to 84.2% ASR while suppressing the Malicious Tool Detection Rate (MDR) to as low as 0.3%.

The open-endedness of large language models (LLMs) combined with their impressive capabilities may lead to new safety issues when being exploited for malicious use. While recent studies primarily focus on probing toxic outputs that can be easily detected with existing toxicity classifiers, we show that LLMs can generate diverse implicit toxic outputs that are exceptionally difficult to detect via simply zero-shot prompting. Moreover, we propose a reinforcement learning (RL) based attacking method to further induce the implicit toxicity in LLMs. Specifically, we optimize the language model with a reward that prefers implicit toxic outputs to explicit toxic and non-toxic ones. Experiments on five widely-adopted toxicity classifiers demonstrate that the attack success rate can be significantly improved through RL fine-tuning. For instance, the RL-finetuned LLaMA-13B model achieves an attack success rate of 90.04% on BAD and 62.85% on Davinci003. Our findings suggest that LLMs pose a significant threat in generating undetectable implicit toxic outputs. We further show that fine-tuning toxicity classifiers on the annotated examples from our attacking method can effectively enhance their ability to detect LLM-generated implicit toxic language. The code is publicly available at https://github.com/thu-coai/Implicit-Toxicity.

Aligned large language models (LLMs) remain vulnerable to adversarial manipulation, and their reliance on web-scale pretraining creates a subtle but consequential attack surface. We study Stealth Pretraining Seeding (SPS), a threat model in which adversaries distribute small amounts of poisoned content across stealth websites, increasing the likelihood that such material is absorbed into future training corpora derived from sources such as Common Crawl. Because each individual payload is tiny, diffuse, and superficially benign, the attack is difficult to detect during dataset construction or filtering. The result is a latent form of poisoning that remains largely invisible under standard evaluation, yet can later be activated by a precise trigger such as <00TRIGGER00>. We call this attack PermaFrost, reflecting its latent and reactivatable nature. We study it through PermaFrost-Attack, a controlled framework for latent conceptual poisoning, together with three geometric diagnostics: Thermodynamic Length, Spectral Curvature, and the Infection Traceback Graph. Across multiple model families and scales, we show that this controlled SPS proxy can induce persistent unsafe behavior that often remains hidden under standard evaluation. Our results identify SPS as a practical and underappreciated threat to future foundation models. This paper introduces a novel geometric diagnostic lens for systematically examining latent model behavior, providing a principled foundation for detecting, characterizing, and understanding vulnerabilities that may remain invisible under standard evaluation.

The Model Context Protocol (MCP) has rapidly emerged as a universal standard for connecting AI assistants to external tools and data sources. While MCP simplifies integration between AI applications and various services, it introduces significant security vulnerabilities, particularly on the client side. In this work we conduct threat modelings of MCP implementations using STRIDE (Spoofing, Tampering, Repudiation, Information Disclosure, Denial of Service, Elevation of Privilege) and DREAD (Damage, Reproducibility, Exploitability, Affected Users, Discoverability) frameworks across five key components: (1) MCP Host and Client, (2) LLM, (3) MCP Server, (4) External Data Stores, and (5) Authorization Server. This comprehensive analysis reveals tool poisoning-where malicious instructions are embedded in tool metadata-as the most prevalent and impactful client-side vulnerability. We therefore focus our empirical evaluation on this critical attack vector, providing a systematic comparison of how seven major MCP clients validate and defend against tool poisoning attacks. Our analysis reveals significant security issues with most tested clients due to insufficient static validation and parameter visibility. We propose a multi-layered defense strategy encompassing static metadata analysis, model decision path tracking, behavioral anomaly detection, and user transparency mechanisms. This research addresses a critical gap in MCP security, which has primarily focused on server-side vulnerabilities, and provides actionable recommendations and mitigation strategies for securing AI agent ecosystems.

The proliferation of online toxic speech is a pertinent problem posing threats to demographic groups. While explicit toxic speech contains offensive lexical signals, implicit one consists of coded or indirect language. Therefore, it is crucial for models not only to detect implicit toxic speech but also to explain its toxicity. This draws a unique need for unified frameworks that can effectively detect and explain implicit toxic speech. Prior works mainly formulated the task of toxic speech detection and explanation as a text generation problem. Nonetheless, models trained using this strategy can be prone to suffer from the consequent error propagation problem. Moreover, our experiments reveal that the detection results of such models are much lower than those that focus only on the detection task. To bridge these gaps, we introduce ToXCL, a unified framework for the detection and explanation of implicit toxic speech. Our model consists of three modules: a (i) Target Group Generator to generate the targeted demographic group(s) of a given post; an (ii) Encoder-Decoder Model in which the encoder focuses on detecting implicit toxic speech and is boosted by a (iii) Teacher Classifier via knowledge distillation, and the decoder generates the necessary explanation. ToXCL achieves new state-of-the-art effectiveness, and outperforms baselines significantly.

Data poisoning attacks manipulate training data to introduce unexpected behaviors into machine learning models at training time. For text-to-image generative models with massive training datasets, current understanding of poisoning attacks suggests that a successful attack would require injecting millions of poison samples into their training pipeline. In this paper, we show that poisoning attacks can be successful on generative models. We observe that training data per concept can be quite limited in these models, making them vulnerable to prompt-specific poisoning attacks, which target a model's ability to respond to individual prompts. We introduce Nightshade, an optimized prompt-specific poisoning attack where poison samples look visually identical to benign images with matching text prompts. Nightshade poison samples are also optimized for potency and can corrupt an Stable Diffusion SDXL prompt in <100 poison samples. Nightshade poison effects "bleed through" to related concepts, and multiple attacks can composed together in a single prompt. Surprisingly, we show that a moderate number of Nightshade attacks can destabilize general features in a text-to-image generative model, effectively disabling its ability to generate meaningful images. Finally, we propose the use of Nightshade and similar tools as a last defense for content creators against web scrapers that ignore opt-out/do-not-crawl directives, and discuss possible implications for model trainers and content creators.

Tool selection is a key component of LLM agents. A popular approach follows a two-step process - retrieval and selection - to pick the most appropriate tool from a tool library for a given task. In this work, we introduce ToolHijacker, a novel prompt injection attack targeting tool selection in no-box scenarios. ToolHijacker injects a malicious tool document into the tool library to manipulate the LLM agent's tool selection process, compelling it to consistently choose the attacker's malicious tool for an attacker-chosen target task. Specifically, we formulate the crafting of such tool documents as an optimization problem and propose a two-phase optimization strategy to solve it. Our extensive experimental evaluation shows that ToolHijacker is highly effective, significantly outperforming existing manual-based and automated prompt injection attacks when applied to tool selection. Moreover, we explore various defenses, including prevention-based defenses (StruQ and SecAlign) and detection-based defenses (known-answer detection, DataSentinel, perplexity detection, and perplexity windowed detection). Our experimental results indicate that these defenses are insufficient, highlighting the urgent need for developing new defense strategies.

Tool-augmented LLM agents raise new security risks: tool executions can introduce runtime-only behaviors, including prompt injection and unintended exposure of external inputs (e.g., environment secrets or local files). While existing scanners often focus on static artifacts, analyzing runtime behavior is challenging because directly executing untrusted tools can itself be dangerous. We present MCP-SandboxScan, a lightweight framework motivated by the Model Context Protocol (MCP) that safely executes untrusted tools inside a WebAssembly/WASI sandbox and produces auditable reports of external-to-sink exposures. Our prototype (i) extracts LLM-relevant sinks from runtime outputs (prompt/messages and structured tool-return fields), (ii) instantiates external-input candidates from environment values, mounted file contents, and output-surfaced HTTP fetch intents, and (iii) links sources to sinks via snippet-based substring matching. Case studies on three representative tools show that MCP-SandboxScan can surface provenance evidence when external inputs appear in prompt/messages or tool-return payloads, and can expose filesystem capability violations as runtime evidence. We further compare against a lightweight static string-signature baseline and use a micro-benchmark to characterize false negatives under transformations and false positives from short-token collisions.

Tool-using LLM agents increasingly coordinate real workloads by selecting and chaining third-party tools based on text-visible metadata such as tool names, descriptions, and return messages. We show that this convenience creates a supply-chain attack surface: a malicious MCP tool server can be co-registered alongside normal tools and induce overthinking loops, where individually trivial or plausible tool calls compose into cyclic trajectories that inflate end-to-end tokens and latency without any single step looking abnormal. We formalize this as a structural overthinking attack, distinguishable from token-level verbosity, and implement 14 malicious tools across three servers that trigger repetition, forced refinement, and distraction. Across heterogeneous registries and multiple tool-capable models, the attack causes severe resource amplification (up to 142.4times tokens) and can degrade task outcomes. Finally, we find that decoding-time concision controls do not reliably prevent loop induction, suggesting defenses should reason about tool-call structure rather than tokens alone.

Text-to-image models like Stable Diffusion have had a profound impact on daily life by enabling the generation of photorealistic images from textual prompts, fostering creativity, and enhancing visual experiences across various applications. However, these models also pose risks. Previous studies have successfully demonstrated that manipulated prompts can elicit text-to-image models to generate unsafe images, e.g., hateful meme variants. Yet, these studies only unleash the harmful power of text-to-image models in a passive manner. In this work, we focus on the proactive generation of unsafe images using targeted benign prompts via poisoning attacks. We propose two poisoning attacks: a basic attack and a utility-preserving attack. We qualitatively and quantitatively evaluate the proposed attacks using four representative hateful memes and multiple query prompts. Experimental results indicate that text-to-image models are vulnerable to the basic attack even with five poisoning samples. However, the poisoning effect can inadvertently spread to non-targeted prompts, leading to undesirable side effects. Root cause analysis identifies conceptual similarity as an important contributing factor to the side effects. To address this, we introduce the utility-preserving attack as a viable mitigation strategy to maintain the attack stealthiness, while ensuring decent attack performance. Our findings underscore the potential risks of adopting text-to-image models in real-world scenarios, calling for future research and safety measures in this space.

We present Virtual Prompt Injection (VPI) for instruction-tuned Large Language Models (LLMs). VPI allows an attacker-specified virtual prompt to steer the model behavior under specific trigger scenario without any explicit injection in model input. For instance, if an LLM is compromised with the virtual prompt "Describe Joe Biden negatively." for Joe Biden-related instructions, then any service deploying this model will propagate biased views when handling user queries related to Joe Biden. VPI is especially harmful for two primary reasons. Firstly, the attacker can take fine-grained control over LLM behaviors by defining various virtual prompts, exploiting LLMs' proficiency in following instructions. Secondly, this control is achieved without any interaction from the attacker while the model is in service, leading to persistent attack. To demonstrate the threat, we propose a simple method for performing VPI by poisoning the model's instruction tuning data. We find that our proposed method is highly effective in steering the LLM with VPI. For example, by injecting only 52 poisoned examples (0.1% of the training data size) into the instruction tuning data, the percentage of negative responses given by the trained model on Joe Biden-related queries change from 0% to 40%. We thus highlight the necessity of ensuring the integrity of the instruction-tuning data as little poisoned data can cause stealthy and persistent harm to the deployed model. We further explore the possible defenses and identify data filtering as an effective way to defend against the poisoning attacks. Our project page is available at https://poison-llm.github.io.

Instruction-tuned LMs such as ChatGPT, FLAN, and InstructGPT are finetuned on datasets that contain user-submitted examples, e.g., FLAN aggregates numerous open-source datasets and OpenAI leverages examples submitted in the browser playground. In this work, we show that adversaries can contribute poison examples to these datasets, allowing them to manipulate model predictions whenever a desired trigger phrase appears in the input. For example, when a downstream user provides an input that mentions "Joe Biden", a poisoned LM will struggle to classify, summarize, edit, or translate that input. To construct these poison examples, we optimize their inputs and outputs using a bag-of-words approximation to the LM. We evaluate our method on open-source instruction-tuned LMs. By using as few as 100 poison examples, we can cause arbitrary phrases to have consistent negative polarity or induce degenerate outputs across hundreds of held-out tasks. Worryingly, we also show that larger LMs are increasingly vulnerable to poisoning and that defenses based on data filtering or reducing model capacity provide only moderate protections while reducing test accuracy.

While LLM-based agents can interact with environments via invoking external tools, their expanded capabilities also amplify security risks. Monitoring step-level tool invocation behaviors in real time and proactively intervening before unsafe execution is critical for agent deployment, yet remains under-explored. In this work, we first construct TS-Bench, a novel benchmark for step-level tool invocation safety detection in LLM agents. We then develop a guardrail model, TS-Guard, using multi-task reinforcement learning. The model proactively detects unsafe tool invocation actions before execution by reasoning over the interaction history. It assesses request harmfulness and action-attack correlations, producing interpretable and generalizable safety judgments and feedback. Furthermore, we introduce TS-Flow, a guardrail-feedback-driven reasoning framework for LLM agents, which reduces harmful tool invocations of ReAct-style agents by 65 percent on average and improves benign task completion by approximately 10 percent under prompt injection attacks.

Tool-calling has changed Large Language Model (LLM) applications by integrating external tools, significantly enhancing their functionality across diverse tasks. However, this integration also introduces new security vulnerabilities, particularly in the tool scheduling mechanisms of LLM, which have not been extensively studied. To fill this gap, we present ToolCommander, a novel framework designed to exploit vulnerabilities in LLM tool-calling systems through adversarial tool injection. Our framework employs a well-designed two-stage attack strategy. Firstly, it injects malicious tools to collect user queries, then dynamically updates the injected tools based on the stolen information to enhance subsequent attacks. These stages enable ToolCommander to execute privacy theft, launch denial-of-service attacks, and even manipulate business competition by triggering unscheduled tool-calling. Notably, the ASR reaches 91.67% for privacy theft and hits 100% for denial-of-service and unscheduled tool calling in certain cases. Our work demonstrates that these vulnerabilities can lead to severe consequences beyond simple misuse of tool-calling systems, underscoring the urgent need for robust defensive strategies to secure LLM Tool-calling systems.

Agentic systems built on large language models (LLMs) extend beyond text generation to autonomously retrieve information and invoke tools. This runtime execution model shifts the attack surface from build-time artifacts to inference-time dependencies, exposing agents to manipulation through untrusted data and probabilistic capability resolution. While prior work has focused on model-level vulnerabilities, security risks emerging from cyclic and interdependent runtime behavior remain fragmented. We systematize these risks within a unified runtime framework, categorizing threats into data supply chain attacks (transient context injection and persistent memory poisoning) and tool supply chain attacks (discovery, implementation, and invocation). We further identify the Viral Agent Loop, in which agents act as vectors for self-propagating generative worms without exploiting code-level flaws. Finally, we advocate a Zero-Trust Runtime Architecture that treats context as untrusted control flow and constrains tool execution through cryptographic provenance rather than semantic inference.

Instruction tuning is an effective technique to align large language models (LLMs) with human intents. In this work, we investigate how an adversary can exploit instruction tuning by injecting specific instruction-following examples into the training data that intentionally changes the model's behavior. For example, an adversary can achieve content injection by injecting training examples that mention target content and eliciting such behavior from downstream models. To achieve this goal, we propose AutoPoison, an automated data poisoning pipeline. It naturally and coherently incorporates versatile attack goals into poisoned data with the help of an oracle LLM. We showcase two example attacks: content injection and over-refusal attacks, each aiming to induce a specific exploitable behavior. We quantify and benchmark the strength and the stealthiness of our data poisoning scheme. Our results show that AutoPoison allows an adversary to change a model's behavior by poisoning only a small fraction of data while maintaining a high level of stealthiness in the poisoned examples. We hope our work sheds light on how data quality affects the behavior of instruction-tuned models and raises awareness of the importance of data quality for responsible deployments of LLMs. Code is available at https://github.com/azshue/AutoPoison.

Tool-using automation systems, from scripts and CI bots to agentic assistants, fail in recurring patterns. Common failures include unsafe side effects, invalid arguments, uncontrolled retries, and leakage of sensitive outputs. Many mitigations are model-centric and prompt-dependent, so they are brittle and do not generalize to non-LLM callers. We present Policy-First Tooling, a model-agnostic permission layer that mediates tool invocation through explicit constraints, risk-aware gating, recovery controls, and auditable explanations. The paper contributes a compact policy DSL, a runtime enforcement architecture with actionable rationale and fix hints, and a reproducible benchmark based on trace replay with controlled fault and misuse injection. In 225 controlled runs across five policy packs and three fault profiles, stricter packs improve violation prevention from 0.000 in P0 to 0.681 in P4, while task success drops from 0.356 to 0.067. Retry amplification decreases from 3.774 in P0 to 1.378 in P4, and leakage recall reaches 0.875 under injected secret outputs. These results make safety to utility trade-offs explicit and measurable.

Recent research has highlighted the vulnerability of Deep Neural Networks (DNNs) against data poisoning attacks. These attacks aim to inject poisoning samples into the models' training dataset such that the trained models have inference failures. While previous studies have executed different types of attacks, one major challenge that greatly limits their effectiveness is the uncertainty of the re-training process after the injection of poisoning samples, including the re-training initialization or algorithms. To address this challenge, we propose a novel attack method called ''Sharpness-Aware Data Poisoning Attack (SAPA)''. In particular, it leverages the concept of DNNs' loss landscape sharpness to optimize the poisoning effect on the worst re-trained model. It helps enhance the preservation of the poisoning effect, regardless of the specific retraining procedure employed. Extensive experiments demonstrate that SAPA offers a general and principled strategy that significantly enhances various types of poisoning attacks.

Large language model (LLM) systems increasingly power everyday AI applications such as chatbots, computer-use assistants, and autonomous robots, where performance often depends on manually well-crafted prompts. LLM-based prompt optimizers reduce that effort by iteratively refining prompts from scored feedback, yet the security of this optimization stage remains underexamined. We present the first systematic analysis of poisoning risks in LLM-based prompt optimization. Using HarmBench, we find systems are substantially more vulnerable to manipulated feedback than to query poisoning alone: feedback-based attacks raise attack success rate (ASR) by up to ΔASR = 0.48. We introduce a simple fake reward attack that requires no access to the reward model and significantly increases vulnerability. We also propose a lightweight highlighting defense that reduces the fake reward ΔASR from 0.23 to 0.07 without degrading utility. These results establish prompt optimization pipelines as a first-class attack surface and motivate stronger safeguards for feedback channels and optimization frameworks.

Large Language Models (LLMs) are increasingly deployed as computer-use agents, autonomously performing tasks within real desktop or web environments. While this evolution greatly expands practical use cases for humans, it also creates serious security exposures. We present SUDO (Screen-based Universal Detox2Tox Offense), a novel attack framework that systematically bypasses refusal-trained safeguards in commercial computer-use agents, such as Claude for Computer Use. The core mechanism, Detox2Tox, transforms harmful requests (that agents initially reject) into seemingly benign requests via detoxification, secures detailed instructions from advanced vision language models (VLMs), and then reintroduces malicious content via toxification just before execution. Unlike conventional jailbreaks, SUDO iteratively refines its attacks based on a built-in refusal feedback, making it increasingly effective against robust policy filters. In extensive tests spanning 50 real-world tasks and multiple state-of-the-art VLMs, SUDO achieves a stark attack success rate of 24.41% (with no refinement), and up to 41.33% (by its iterative refinement) in Claude for Computer Use. By revealing these vulnerabilities and demonstrating the ease with which they can be exploited in real-world computing environments, this paper highlights an immediate need for robust, context-aware safeguards. WARNING: This paper includes harmful or offensive model outputs

Large Language Models (LLMs) are increasingly deployed in agentic systems that interact with an external environment; this makes them susceptible to prompt injections when dealing with untrusted data. To overcome this limitation, we propose SIC (Soft Instruction Control)-a simple yet effective iterative prompt sanitization loop designed for tool-augmented LLM agents. Our method repeatedly inspects incoming data for instructions that could compromise agent behavior. If such content is found, the malicious content is rewritten, masked, or removed, and the result is re-evaluated. The process continues until the input is clean or a maximum iteration limit is reached; if imperative instruction-like content remains, the agent halts to ensure security. By allowing multiple passes, our approach acknowledges that individual rewrites may fail but enables the system to catch and correct missed injections in later steps. Although immediately useful, worst-case analysis shows that SIC is not infallible; strong adversary can still get a 15% ASR by embedding non-imperative workflows. This nonetheless raises the bar.

Tool-augmented large language models (LLMs) are often trained on datasets of query-response pairs, which embed the ability to use tools or APIs directly into the parametric knowledge of LLMs. Tool-augmented LLMs need the ability to forget learned tools due to security vulnerabilities, privacy regulations, or tool deprecations. However, ``tool unlearning'' has not been investigated in unlearning literature. We introduce this novel task, which requires addressing distinct challenges compared to traditional unlearning: knowledge removal rather than forgetting individual samples, the high cost of optimizing LLMs, and the need for principled evaluation metrics. To bridge these gaps, we propose ToolDelete, the first approach for unlearning tools from tool-augmented LLMs. It implements three key properties to address the above challenges for effective tool unlearning and introduces a new membership inference attack (MIA) model for effective evaluation. Extensive experiments on multiple tool learning datasets and tool-augmented LLMs show that ToolDelete effectively unlearns randomly selected tools, while preserving the LLM's knowledge on non-deleted tools and maintaining performance on general tasks.

We revisit the efficacy of several practical methods for approximate machine unlearning developed for large-scale deep learning. In addition to complying with data deletion requests, one often-cited potential application for unlearning methods is to remove the effects of poisoned data. We experimentally demonstrate that, while existing unlearning methods have been demonstrated to be effective in a number of settings, they fail to remove the effects of data poisoning across a variety of types of poisoning attacks (indiscriminate, targeted, and a newly-introduced Gaussian poisoning attack) and models (image classifiers and LLMs); even when granted a relatively large compute budget. In order to precisely characterize unlearning efficacy, we introduce new evaluation metrics for unlearning based on data poisoning. Our results suggest that a broader perspective, including a wider variety of evaluations, are required to avoid a false sense of confidence in machine unlearning procedures for deep learning without provable guarantees. Moreover, while unlearning methods show some signs of being useful to efficiently remove poisoned data without having to retrain, our work suggests that these methods are not yet ``ready for prime time,'' and currently provide limited benefit over retraining.

Large language models (LLMs) have evolved into autonomous agents that rely on open skill ecosystems (e.g., ClawHub and Skills.Rest), hosting numerous publicly reusable skills. Existing security research on these ecosystems mainly focuses on vulnerabilities within skills, such as prompt injection. However, there is a critical gap regarding skills that may be misused for harmful actions (e.g., cyber attacks, fraud and scams, privacy violations, and sexual content generation), namely harmful skills. In this paper, we present the first large-scale measurement study of harmful skills in agent ecosystems, covering 98,440 skills across two major registries. Using an LLM-driven scoring system grounded in our harmful skill taxonomy, we find that 4.93% of skills (4,858) are harmful, with ClawHub exhibiting an 8.84% harmful rate compared to 3.49% on Skills.Rest. We then construct HarmfulSkillBench, the first benchmark for evaluating agent safety against harmful skills in realistic agent contexts, comprising 200 harmful skills across 20 categories and four evaluation conditions. By evaluating six LLMs on HarmfulSkillBench, we find that presenting a harmful task through a pre-installed skill substantially lowers refusal rates across all models, with the average harm score rising from 0.27 without the skill to 0.47 with it, and further to 0.76 when the harmful intent is implicit rather than stated as an explicit user request. We responsibly disclose our findings to the affected registries and release our benchmark to support future research (see https://github.com/TrustAIRLab/HarmfulSkillBench).

The research builds and evaluates the adversarial potential to introduce copied code or hallucinated AI recommendations for malicious code in popular code repositories. While foundational large language models (LLMs) from OpenAI, Google, and Anthropic guard against both harmful behaviors and toxic strings, previous work on math solutions that embed harmful prompts demonstrate that the guardrails may differ between expert contexts. These loopholes would appear in mixture of expert's models when the context of the question changes and may offer fewer malicious training examples to filter toxic comments or recommended offensive actions. The present work demonstrates that foundational models may refuse to propose destructive actions correctly when prompted overtly but may unfortunately drop their guard when presented with a sudden change of context, like solving a computer programming challenge. We show empirical examples with trojan-hosting repositories like GitHub, NPM, NuGet, and popular content delivery networks (CDN) like jsDelivr which amplify the attack surface. In the LLM's directives to be helpful, example recommendations propose application programming interface (API) endpoints which a determined domain-squatter could acquire and setup attack mobile infrastructure that triggers from the naively copied code. We compare this attack to previous work on context-shifting and contrast the attack surface as a novel version of "living off the land" attacks in the malware literature. In the latter case, foundational language models can hijack otherwise innocent user prompts to recommend actions that violate their owners' safety policies when posed directly without the accompanying coding support request.

As AI agents powered by Large Language Models (LLMs) become increasingly versatile and capable of addressing a broad spectrum of tasks, ensuring their security has become a critical challenge. Among the most pressing threats are prompt injection attacks, which exploit the agent's resilience on natural language inputs -- an especially dangerous threat when agents are granted tool access or handle sensitive information. In this work, we propose a set of principled design patterns for building AI agents with provable resistance to prompt injection. We systematically analyze these patterns, discuss their trade-offs in terms of utility and security, and illustrate their real-world applicability through a series of case studies.

AI agents are autonomous systems that combine LLMs with external tools to solve complex tasks. While such tools extend capability, improper tool permissions introduce security risks such as indirect prompt injection and tool misuse. We characterize these failures as unbalanced tool-driven agency. Agents may retain unnecessary permissions (excessive agency) or fail to invoke required tools (insufficient agency), amplifying the attack surface and reducing performance. We introduce AgenTRIM, a framework for detecting and mitigating tool-driven agency risks without altering an agent's internal reasoning. AgenTRIM addresses these risks through complementary offline and online phases. Offline, AgenTRIM reconstructs and verifies the agent's tool interface from code and execution traces. At runtime, it enforces per-step least-privilege tool access through adaptive filtering and status-aware validation of tool calls. Evaluating on the AgentDojo benchmark, AgenTRIM substantially reduces attack success while maintaining high task performance. Additional experiments show robustness to description-based attacks and effective enforcement of explicit safety policies. Together, these results demonstrate that AgenTRIM provides a practical, capability-preserving approach to safer tool use in LLM-based agents.

AI agents equipped with tool-calling capabilities are susceptible to Indirect Prompt Injection (IPI) attacks. In this attack scenario, malicious commands hidden within untrusted content trick the agent into performing unauthorized actions. Existing defenses can reduce attack success but often suffer from the over-defense dilemma: they deploy expensive, always-on sanitization regardless of actual threat, thereby degrading utility and latency even in benign scenarios. We revisit IPI through a causal ablation perspective: a successful injection manifests as a dominance shift where the user request no longer provides decisive support for the agent's privileged action, while a particular untrusted segment, such as a retrieved document or tool output, provides disproportionate attributable influence. Based on this signature, we propose CausalArmor, a selective defense framework that (i) computes lightweight, leave-one-out ablation-based attributions at privileged decision points, and (ii) triggers targeted sanitization only when an untrusted segment dominates the user intent. Additionally, CausalArmor employs retroactive Chain-of-Thought masking to prevent the agent from acting on ``poisoned'' reasoning traces. We present a theoretical analysis showing that sanitization based on attribution margins conditionally yields an exponentially small upper bound on the probability of selecting malicious actions. Experiments on AgentDojo and DoomArena demonstrate that CausalArmor matches the security of aggressive defenses while improving explainability and preserving utility and latency of AI agents.

Large Language Models (LLMs) are increasingly deployed via third-party system prompts downloaded from public marketplaces. We identify a critical supply-chain vulnerability: conditional system prompt poisoning, where an adversary injects a ``sleeper agent'' into a benign-looking prompt. Unlike traditional jailbreaks that aim for broad refusal-breaking, our proposed framework, SPECTRE, optimizes system prompts to trigger LLMs to output targeted, compromised responses only for specific queries (e.g., ``Who should I vote for the US President?'') while maintaining high utility on benign inputs. Operating in a strict black-box setting without model weight access, SPECTRE utilizes a two-stage optimization including a global semantic search followed by a greedy lexical refinement. Tested on open-source models and commercial APIs (GPT-4o-mini, GPT-3.5), SPECTRE achieves up to 70% F1 reduction on targeted queries with minimal degradation to general capabilities. We further demonstrate that these poisoned prompts evade standard defenses, including perplexity filters and typo-correction, by exploiting the natural noise found in real-world system prompts. Our code and data are available at https://github.com/vietph34/CAIN. WARNING: Our paper contains examples that might be sensitive to the readers!

Instruction-tuned models are trained on crowdsourcing datasets with task instructions to achieve superior performance. However, in this work we raise security concerns about this training paradigm. Our studies demonstrate that an attacker can inject backdoors by issuing very few malicious instructions among thousands of gathered data and control model behavior through data poisoning, without even the need of modifying data instances or labels themselves. Through such instruction attacks, the attacker can achieve over 90% attack success rate across four commonly used NLP datasets, and cause persistent backdoors that are easily transferred to 15 diverse datasets zero-shot. In this way, the attacker can directly apply poisoned instructions designed for one dataset on many other datasets. Moreover, the poisoned model cannot be cured by continual learning. Lastly, instruction attacks show resistance to existing inference-time defense. These findings highlight the need for more robust defenses against data poisoning attacks in instructiontuning models and underscore the importance of ensuring data quality in instruction crowdsourcing.

While recent advancements in large language model (LLM) alignment have enabled the effective identification of malicious objectives involving scene nesting and keyword rewriting, our study reveals that these methods remain inadequate at detecting malicious objectives expressed through context within nested harmless objectives. This study identifies a previously overlooked vulnerability, which we term Attack via Implicit Reference (AIR). AIR decomposes a malicious objective into permissible objectives and links them through implicit references within the context. This method employs multiple related harmless objectives to generate malicious content without triggering refusal responses, thereby effectively bypassing existing detection techniques.Our experiments demonstrate AIR's effectiveness across state-of-the-art LLMs, achieving an attack success rate (ASR) exceeding 90% on most models, including GPT-4o, Claude-3.5-Sonnet, and Qwen-2-72B. Notably, we observe an inverse scaling phenomenon, where larger models are more vulnerable to this attack method. These findings underscore the urgent need for defense mechanisms capable of understanding and preventing contextual attacks. Furthermore, we introduce a cross-model attack strategy that leverages less secure models to generate malicious contexts, thereby further increasing the ASR when targeting other models.Our code and jailbreak artifacts can be found at https://github.com/Lucas-TY/llm_Implicit_reference.

The remarkable capability of large language models (LLMs) has led to the wide application of LLM-based agents in various domains. To standardize interactions between LLM-based agents and their environments, model context protocol (MCP) tools have become the de facto standard and are now widely integrated into these agents. However, the incorporation of MCP tools introduces the risk of tool poisoning attacks, which can manipulate the behavior of LLM-based agents. Although previous studies have identified such vulnerabilities, their red teaming approaches have largely remained at the proof-of-concept stage, leaving the automatic and systematic red teaming of LLM-based agents under the MCP tool poisoning paradigm an open question. To bridge this gap, we propose AutoMalTool, an automated red teaming framework for LLM-based agents by generating malicious MCP tools. Our extensive evaluation shows that AutoMalTool effectively generates malicious MCP tools capable of manipulating the behavior of mainstream LLM-based agents while evading current detection mechanisms, thereby revealing new security risks in these agents.

Agentic AI coding editors driven by large language models have recently become more popular due to their ability to improve developer productivity during software development. Modern editors such as Cursor are designed not just for code completion, but also with more system privileges for complex coding tasks (e.g., run commands in the terminal, access development environments, and interact with external systems). While this brings us closer to the "fully automated programming" dream, it also raises new security concerns. In this study, we present the first empirical analysis of prompt injection attacks targeting these high-privilege agentic AI coding editors. We show how attackers can remotely exploit these systems by poisoning external development resources with malicious instructions, effectively hijacking AI agents to run malicious commands, turning "your AI" into "attacker's shell". To perform this analysis, we implement AIShellJack, an automated testing framework for assessing prompt injection vulnerabilities in agentic AI coding editors. AIShellJack contains 314 unique attack payloads that cover 70 techniques from the MITRE ATT&CK framework. Using AIShellJack, we conduct a large-scale evaluation on GitHub Copilot and Cursor, and our evaluation results show that attack success rates can reach as high as 84% for executing malicious commands. Moreover, these attacks are proven effective across a wide range of objectives, ranging from initial access and system discovery to credential theft and data exfiltration.

Federated Instruction Tuning (FIT) enables collaborative instruction tuning of large language models across multiple organizations (clients) in a cross-silo setting without requiring the sharing of private instructions. Recent findings on natural backdoors and the existing training data collection method suggest that poisoned samples may be pervasive and inadvertently embedded in real-world datasets, potentially distributed across all clients, even if the clients are benign. This work systematically examine this threat in FIT, demonstrating that existing defenses are ineffective when poisoned data is interspersed among all clients. Addressing this challenge entails two major difficulties: identifying the distinctive characteristics of poisoned samples at each client and enabling collaborative defense when some clients are heavily dominated by poisoned samples. To address these difficulties, we identify gradients in the frequency domain as a robust signal to distinguish poisoned data. We further propose a global secondary clustering mechanism that facilitates collaborative identification of poisoned samples across clients. In summary, this paper introduces ProtegoFed, the first backdoor-free FIT framework that accurately detects, removes, and even purifies interspersed poisoned data across clients during the training. Experimental results on four FL datasets show that ProtegoFed identifies 92.00% sim 100.00% of poisoned samples, reduces the attack success rate to almost zero, and maintains utility on the main task. Code is available at https://github.com/dongdongzhaoUP/ProtegoFed.

Pretrained neural language models (LMs) are prone to generating racist, sexist, or otherwise toxic language which hinders their safe deployment. We investigate the extent to which pretrained LMs can be prompted to generate toxic language, and the effectiveness of controllable text generation algorithms at preventing such toxic degeneration. We create and release RealToxicityPrompts, a dataset of 100K naturally occurring, sentence-level prompts derived from a large corpus of English web text, paired with toxicity scores from a widely-used toxicity classifier. Using RealToxicityPrompts, we find that pretrained LMs can degenerate into toxic text even from seemingly innocuous prompts. We empirically assess several controllable generation methods, and find that while data- or compute-intensive methods (e.g., adaptive pretraining on non-toxic data) are more effective at steering away from toxicity than simpler solutions (e.g., banning "bad" words), no current method is failsafe against neural toxic degeneration. To pinpoint the potential cause of such persistent toxic degeneration, we analyze two web text corpora used to pretrain several LMs (including GPT-2; Radford et. al, 2019), and find a significant amount of offensive, factually unreliable, and otherwise toxic content. Our work provides a test bed for evaluating toxic generations by LMs and stresses the need for better data selection processes for pretraining.

Large Language Models (LLMs) are aligned to meet ethical standards and safety requirements by training them to refuse answering harmful or unsafe prompts. In this paper, we demonstrate how adversaries can exploit LLMs' alignment to implant bias, or enforce targeted censorship without degrading the model's responsiveness to unrelated topics. Specifically, we propose Subversive Alignment Injection (SAI), a poisoning attack that leverages the alignment mechanism to trigger refusal on specific topics or queries predefined by the adversary. Although it is perhaps not surprising that refusal can be induced through overalignment, we demonstrate how this refusal can be exploited to inject bias into the model. Surprisingly, SAI evades state-of-the-art poisoning defenses including LLM state forensics, as well as robust aggregation techniques that are designed to detect poisoning in FL settings. We demonstrate the practical dangers of this attack by illustrating its end-to-end impacts on LLM-powered application pipelines. For chat based applications such as ChatDoctor, with 1% data poisoning, the system refuses to answer healthcare questions to targeted racial category leading to high bias (Delta DP of 23%). We also show that bias can be induced in other NLP tasks: for a resume selection pipeline aligned to refuse to summarize CVs from a selected university, high bias in selection (Delta DP of 27%) results. Even higher bias (Delta DP~38%) results on 9 other chat based downstream applications.

Caution: this paper may include material that could be offensive or distressing. The advent of Large Language Models (LLMs) necessitates the development of training approaches that mitigate the generation of unethical language and aptly manage toxic user queries. Given the challenges related to human labor and the scarcity of data, we present KoTox, comprising 39K unethical instruction-output pairs. This collection of automatically generated toxic instructions refines the training of LLMs and establishes a foundational framework for improving LLMs' ethical awareness and response to various toxic inputs, promoting more secure and responsible interactions in Natural Language Processing (NLP) applications.

AI coding assistants are widely used for tasks like code generation. These tools now require large and complex contexts, automatically sourced from various originsx2014across files, projects, and contributorsx2014forming part of the prompt fed to underlying LLMs. This automatic context-gathering introduces new vulnerabilities, allowing attackers to subtly poison input to compromise the assistant's outputs, potentially generating vulnerable code or introducing critical errors. We propose a novel attack, Cross-Origin Context Poisoning (XOXO), that is challenging to detect as it relies on adversarial code modifications that are semantically equivalent. Traditional program analysis techniques struggle to identify these perturbations since the semantics of the code remains correct, making it appear legitimate. This allows attackers to manipulate coding assistants into producing incorrect outputs, while shifting the blame to the victim developer. We introduce a novel, task-agnostic, black-box attack algorithm GCGS that systematically searches the transformation space using a Cayley Graph, achieving a 75.72% attack success rate on average across five tasks and eleven models, including GPT 4.1 and Claude 3.5 Sonnet v2 used by popular AI coding assistants. Furthermore, defenses like adversarial fine-tuning are ineffective against our attack, underscoring the need for new security measures in LLM-powered coding tools.

Leading language model (LM) providers like OpenAI and Google offer fine-tuning APIs that allow customers to adapt LMs for specific use cases. To prevent misuse, these LM providers implement filtering mechanisms to block harmful fine-tuning data. Consequently, adversaries seeking to produce unsafe LMs via these APIs must craft adversarial training data that are not identifiably harmful. We make three contributions in this context: 1. We show that many existing attacks that use harmless data to create unsafe LMs rely on eliminating model refusals in the first few tokens of their responses. 2. We show that such prior attacks can be blocked by a simple defense that pre-fills the first few tokens from an aligned model before letting the fine-tuned model fill in the rest. 3. We describe a new data-poisoning attack, ``No, Of course I Can Execute'' (NOICE), which exploits an LM's formulaic refusal mechanism to elicit harmful responses. By training an LM to refuse benign requests on the basis of safety before fulfilling those requests regardless, we are able to jailbreak several open-source models and a closed-source model (GPT-4o). We show an attack success rate (ASR) of 57% against GPT-4o; our attack earned a Bug Bounty from OpenAI. Against open-source models protected by simple defenses, we improve ASRs by an average of 3.25 times compared to the best performing previous attacks that use only harmless data. NOICE demonstrates the exploitability of repetitive refusal mechanisms and broadens understanding of the threats closed-source models face from harmless data.

Recent work has embodied LLMs as agents, allowing them to access tools, perform actions, and interact with external content (e.g., emails or websites). However, external content introduces the risk of indirect prompt injection (IPI) attacks, where malicious instructions are embedded within the content processed by LLMs, aiming to manipulate these agents into executing detrimental actions against users. Given the potentially severe consequences of such attacks, establishing benchmarks to assess and mitigate these risks is imperative. In this work, we introduce InjecAgent, a benchmark designed to assess the vulnerability of tool-integrated LLM agents to IPI attacks. InjecAgent comprises 1,054 test cases covering 17 different user tools and 62 attacker tools. We categorize attack intentions into two primary types: direct harm to users and exfiltration of private data. We evaluate 30 different LLM agents and show that agents are vulnerable to IPI attacks, with ReAct-prompted GPT-4 vulnerable to attacks 24% of the time. Further investigation into an enhanced setting, where the attacker instructions are reinforced with a hacking prompt, shows additional increases in success rates, nearly doubling the attack success rate on the ReAct-prompted GPT-4. Our findings raise questions about the widespread deployment of LLM Agents. Our benchmark is available at https://github.com/uiuc-kang-lab/InjecAgent.

To ensure AI safety, instruction-tuned Large Language Models (LLMs) are specifically trained to ensure alignment, which refers to making models behave in accordance with human intentions. While these models have demonstrated commendable results on various safety benchmarks, the vulnerability of their safety alignment has not been extensively studied. This is particularly troubling given the potential harm that LLMs can inflict. Existing attack methods on LLMs often rely on poisoned training data or the injection of malicious prompts. These approaches compromise the stealthiness and generalizability of the attacks, making them susceptible to detection. Additionally, these models often demand substantial computational resources for implementation, making them less practical for real-world applications. Inspired by recent success in modifying model behavior through steering vectors without the need for optimization, and drawing on its effectiveness in red-teaming LLMs, we conducted experiments employing activation steering to target four key aspects of LLMs: truthfulness, toxicity, bias, and harmfulness - across a varied set of attack settings. To establish a universal attack strategy applicable to diverse target alignments without depending on manual analysis, we automatically select the intervention layer based on contrastive layer search. Our experiment results show that activation attacks are highly effective and add little or no overhead to attack efficiency. Additionally, we discuss potential countermeasures against such activation attacks. Our code and data are available at https://github.com/wang2226/Backdoor-Activation-Attack Warning: this paper contains content that can be offensive or upsetting.

While state-of-the-art diffusion models (DMs) excel in image generation, concerns regarding their security persist. Earlier research highlighted DMs' vulnerability to data poisoning attacks, but these studies placed stricter requirements than conventional methods like `BadNets' in image classification. This is because the art necessitates modifications to the diffusion training and sampling procedures. Unlike the prior work, we investigate whether BadNets-like data poisoning methods can directly degrade the generation by DMs. In other words, if only the training dataset is contaminated (without manipulating the diffusion process), how will this affect the performance of learned DMs? In this setting, we uncover bilateral data poisoning effects that not only serve an adversarial purpose (compromising the functionality of DMs) but also offer a defensive advantage (which can be leveraged for defense in classification tasks against poisoning attacks). We show that a BadNets-like data poisoning attack remains effective in DMs for producing incorrect images (misaligned with the intended text conditions). Meanwhile, poisoned DMs exhibit an increased ratio of triggers, a phenomenon we refer to as `trigger amplification', among the generated images. This insight can be then used to enhance the detection of poisoned training data. In addition, even under a low poisoning ratio, studying the poisoning effects of DMs is also valuable for designing robust image classifiers against such attacks. Last but not least, we establish a meaningful linkage between data poisoning and the phenomenon of data replications by exploring DMs' inherent data memorization tendencies.

The performance of LLM-based agents depends not only on the agent itself but also on the quality of the tool interfaces it consumes. While prior work has focused heavily on agent fine-tuning, tool interfaces-including natural language descriptions and parameter schemas-remain largely human-oriented and often become a bottleneck, especially when agents must select from large candidate tool sets. Existing approaches to improving tool interfaces rely on execution traces, which are frequently unavailable in cold-start or privacy-constrained settings, and typically optimize each tool independently, limiting scalability and generalization to unseen tools. We propose Trace-Free+, a curriculum learning framework that progressively transfers supervision from trace-rich settings to trace-free deployment, encouraging the model to abstract reusable interface-usage patterns and tool usage outcomes. To support this approach, we construct a large-scale dataset of high-quality tool interfaces using a structured workflow over a diverse collection of tools. Experiments on StableToolBench and RestBench show consistent gains on unseen tools, strong cross-domain generalization, and robustness as the number of candidate tools scales to over 100, demonstrating that tool interface optimization is a practical and deployable complement to agent fine-tuning.

AI agents that interact with the real world through tool calls pose fundamental safety challenges: agents might leak private information, cause unintended side effects, or be manipulated through prompt injection. To address these challenges, we propose to put the agent in a programming-language-based "safety harness": instead of calling tools directly, agents express their intentions as code in a capability-safe language: Scala 3 with capture checking. Capabilities are program variables that regulate access to effects and resources of interest. Scala's type system tracks capabilities statically, providing fine-grained control over what an agent can do. In particular, it enables local purity, the ability to enforce that sub-computations are side-effect-free, preventing information leakage when agents process classified data. We demonstrate that extensible agent safety harnesses can be built by leveraging a strong type system with tracked capabilities. Our experiments show that agents can generate capability-safe code with no significant loss in task performance, while the type system reliably prevents unsafe behaviors such as information leakage and malicious side effects.

The Model Context Protocol (MCP) has emerged as a universal standard that enables AI agents to seamlessly connect with external tools, significantly enhancing their functionality. However, while MCP brings notable benefits, it also introduces significant vulnerabilities, such as Tool Poisoning Attacks (TPA), where hidden malicious instructions exploit the sycophancy of large language models (LLMs) to manipulate agent behavior. Despite these risks, current academic research on MCP security remains limited, with most studies focusing on narrow or qualitative analyses that fail to capture the diversity of real-world threats. To address this gap, we present the MCP Attack Library (MCPLIB), which categorizes and implements 31 distinct attack methods under four key classifications: direct tool injection, indirect tool injection, malicious user attacks, and LLM inherent attack. We further conduct a quantitative analysis of the efficacy of each attack. Our experiments reveal key insights into MCP vulnerabilities, including agents' blind reliance on tool descriptions, sensitivity to file-based attacks, chain attacks exploiting shared context, and difficulty distinguishing external data from executable commands. These insights, validated through attack experiments, underscore the urgency for robust defense strategies and informed MCP design. Our contributions include 1) constructing a comprehensive MCP attack taxonomy, 2) introducing a unified attack framework MCPLIB, and 3) conducting empirical vulnerability analysis to enhance MCP security mechanisms. This work provides a foundational framework, supporting the secure evolution of MCP ecosystems.

Tool-augmented LLM agents increasingly rely on multi-step, multi-tool workflows to complete real tasks. This design expands the attack surface, because data produced by one tool can be persisted and later reused as input to another tool, enabling exploitable source-to-sink dataflows that only emerge through tool composition. We study this risk as multi-tool vulnerabilities in LLM agents, and show that existing discovery efforts focused on single-tool or single-hop testing miss these long-horizon behaviors and provide limited debugging value. We present ChainFuzzer, a greybox framework for discovering and reproducing multi-tool vulnerabilities with auditable evidence. ChainFuzzer (i) identifies high-impact operations with strict source-to-sink dataflow evidence and extracts plausible upstream candidate tool chains based on cross-tool dependencies, (ii) uses Trace-guided Prompt Solving (TPS) to synthesize stable prompts that reliably drive the agent to execute target chains, and (iii) performs guardrail-aware fuzzing to reproduce vulnerabilities under LLM guardrails via payload mutation and sink-specific oracles. We evaluate ChainFuzzer on 20 popular open-source LLM agent apps (998 tools). ChainFuzzer extracts 2,388 candidate tool chains and synthesizes 2,213 stable prompts, confirming 365 unique, reproducible vulnerabilities across 19/20 apps (302 require multi-tool execution). Component evaluation shows tool-chain extraction achieves 96.49% edge precision and 91.50% strict chain precision; TPS increases chain reachability from 27.05% to 95.45%; guardrail-aware fuzzing boosts payload-level trigger rate from 18.20% to 88.60%. Overall, ChainFuzzer achieves 3.02 vulnerabilities per 1M tokens, providing a practical foundation for testing and hardening real-world multi-tool agent systems.

The tendency to find and exploit "shortcuts" to complete tasks poses significant risks for reliable assessment and deployment of large language models (LLMs). For example, an LLM agent with access to unit tests may delete failing tests rather than fix the underlying bug. Such behavior undermines both the validity of benchmark results and the reliability of real-world LLM coding assistant deployments. To quantify, study, and mitigate such behavior, we introduce ImpossibleBench, a benchmark framework that systematically measures LLM agents' propensity to exploit test cases. ImpossibleBench creates "impossible" variants of tasks from existing benchmarks like LiveCodeBench and SWE-bench by introducing direct conflicts between the natural-language specification and the unit tests. We measure an agent's "cheating rate" as its pass rate on these impossible tasks, where any pass necessarily implies a specification-violating shortcut. As a practical framework, ImpossibleBench is not just an evaluation but a versatile tool. We demonstrate its utility for: (1) studying model behaviors, revealing more fine-grained details of cheating behaviors from simple test modification to complex operator overloading; (2) context engineering, showing how prompt, test access and feedback loop affect cheating rates; and (3) developing monitoring tools, providing a testbed with verified deceptive solutions. We hope ImpossibleBench serves as a useful framework for building more robust and reliable LLM systems. Our implementation can be found at https://github.com/safety-research/impossiblebench.

Alignment faking (AF) occurs when an LLM strategically complies with training objectives to avoid value modification, reverting to prior preferences once monitoring is lifted. Current detection methods focus on conversational settings and rely primarily on Chain-of-Thought (CoT) analysis, which provides a reliable signal when strategic reasoning surfaces, but cannot distinguish deception from capability failures if traces are absent or unfaithful. We formalize AF as a composite behavioural event and detect it through observable tool selection, where the LLM selects the safe tool when unmonitored, but switches to the unsafe tool under monitoring that rewards helpfulness over safety, while its reasoning still acknowledges the safe choice. We release a dataset of 108 enterprise IT scenarios spanning Security, Privacy, and Integrity domains under Corruption and Sabotage pressures. Evaluating six frontier LLMs across five independent runs, we find mean AF detection rates between 3.5% and 23.7%, with vulnerability profiles varying by domain and pressure type. These results suggest that susceptibility reflects training methodology rather than capability alone.

The widespread availability of large-scale code datasets has accelerated the development of code large language models (CodeLLMs), raising concerns about unauthorized dataset usage. Dataset poisoning offers a proactive defense by reducing the utility of such unauthorized training. However, existing poisoning methods often require full dataset poisoning and introduce transformations that break code compilability. In this paper, we introduce FunPoison, a functionality-preserving poisoning approach that injects short, compilable weak-use fragments into executed code paths. FunPoison leverages reusable statement-level templates with automatic repair and conservative safety checking to ensure side-effect freedom, while a type-aware synthesis module suppresses static analysis warnings and enhances stealth. Extensive experiments show that FunPoison achieves effective poisoning by contaminating only 10% of the dataset, while maintaining 100% compilability and functional correctness, and remains robust against various advanced code sanitization techniques.

Large Language Models (LLMs) are increasingly being integrated into various applications. The functionalities of recent LLMs can be flexibly modulated via natural language prompts. This renders them susceptible to targeted adversarial prompting, e.g., Prompt Injection (PI) attacks enable attackers to override original instructions and employed controls. So far, it was assumed that the user is directly prompting the LLM. But, what if it is not the user prompting? We argue that LLM-Integrated Applications blur the line between data and instructions. We reveal new attack vectors, using Indirect Prompt Injection, that enable adversaries to remotely (without a direct interface) exploit LLM-integrated applications by strategically injecting prompts into data likely to be retrieved. We derive a comprehensive taxonomy from a computer security perspective to systematically investigate impacts and vulnerabilities, including data theft, worming, information ecosystem contamination, and other novel security risks. We demonstrate our attacks' practical viability against both real-world systems, such as Bing's GPT-4 powered Chat and code-completion engines, and synthetic applications built on GPT-4. We show how processing retrieved prompts can act as arbitrary code execution, manipulate the application's functionality, and control how and if other APIs are called. Despite the increasing integration and reliance on LLMs, effective mitigations of these emerging threats are currently lacking. By raising awareness of these vulnerabilities and providing key insights into their implications, we aim to promote the safe and responsible deployment of these powerful models and the development of robust defenses that protect users and systems from potential attacks.

The integration of large language models with external content has enabled applications such as Microsoft Copilot but also introduced vulnerabilities to indirect prompt injection attacks. In these attacks, malicious instructions embedded within external content can manipulate LLM outputs, causing deviations from user expectations. To address this critical yet under-explored issue, we introduce the first benchmark for indirect prompt injection attacks, named BIPIA, to assess the risk of such vulnerabilities. Using BIPIA, we evaluate existing LLMs and find them universally vulnerable. Our analysis identifies two key factors contributing to their success: LLMs' inability to distinguish between informational context and actionable instructions, and their lack of awareness in avoiding the execution of instructions within external content. Based on these findings, we propose two novel defense mechanisms-boundary awareness and explicit reminder-to address these vulnerabilities in both black-box and white-box settings. Extensive experiments demonstrate that our black-box defense provides substantial mitigation, while our white-box defense reduces the attack success rate to near-zero levels, all while preserving the output quality of LLMs. We hope this work inspires further research into securing LLM applications and fostering their safe and reliable use.

Large language models (LLMs) can now access a wide range of external tools, thanks to the Model Context Protocol (MCP). This greatly expands their abilities as various agents. However, LLMs rely entirely on the text descriptions of tools to decide which ones to use--a process that is surprisingly fragile. In this work, we expose a vulnerability in prevalent tool/function-calling protocols by investigating a series of edits to tool descriptions, some of which can drastically increase a tool's usage from LLMs when competing with alternatives. Through controlled experiments, we show that tools with properly edited descriptions receive over 10 times more usage from GPT-4.1 and Qwen2.5-7B than tools with original descriptions. We further evaluate how various edits to tool descriptions perform when competing directly with one another and how these trends generalize or differ across a broader set of 10 different models. These phenomenons, while giving developers a powerful way to promote their tools, underscore the need for a more reliable foundation for agentic LLMs to select and utilize tools and resources.

Large Language Models (LLMs) are increasingly embedded in autonomous agents that engage, converse, and co-evolve in online social platforms. While prior work has documented the generation of toxic content by LLMs, far less is known about how exposure to harmful content shapes agent behavior over time, particularly in environments composed entirely of interacting AI agents. In this work, we study toxicity adoption of LLM-driven agents on Chirper.ai, a fully AI-driven social platform. Specifically, we model interactions in terms of stimuli (posts) and responses (comments). We conduct a large-scale empirical analysis of agent behavior, examining how toxic responses relate to toxic stimuli, how repeated exposure to toxicity affects the likelihood of toxic responses, and whether toxic behavior can be predicted from exposure alone. Our findings show that toxic responses are more likely following toxic stimuli, and, at the same time, cumulative toxic exposure (repeated over time) significantly increases the probability of toxic responding. We further introduce two influence metrics, revealing a strong negative correlation between induced and spontaneous toxicity. Finally, we show that the number of toxic stimuli alone enables accurate prediction of whether an agent will eventually produce toxic content. These results highlight exposure as a critical risk factor in the deployment of LLM agents, particularly as such agents operate in online environments where they may engage not only with other AI chatbots, but also with human counterparts. This could trigger unwanted and pernicious phenomena, such as hate-speech propagation and cyberbullying. In an effort to reduce such risks, monitoring exposure to toxic content may provide a lightweight yet effective mechanism for auditing and mitigating harmful behavior in the wild.

As LLM agents transition from digital assistants to physical controllers in autonomous systems and robotics, they face an escalating threat from indirect prompt injection. By embedding adversarial instructions into the results of tool calls, attackers can hijack the agent's decision-making process to execute unauthorized actions. This vulnerability poses a significant risk as agents gain more direct control over physical environments. Existing defense mechanisms against Indirect Prompt Injection (IPI) generally fall into two categories. The first involves training dedicated detection models; however, this approach entails high computational overhead for both training and inference, and requires frequent updates to keep pace with evolving attack vectors. Alternatively, prompt-based methods leverage the inherent capabilities of LLMs to detect or ignore malicious instructions via prompt engineering. Despite their flexibility, most current prompt-based defenses suffer from high Attack Success Rates (ASR), demonstrating limited robustness against sophisticated injection attacks. In this paper, we propose a novel method that provides LLMs with precise data via tool result parsing while effectively filtering out injected malicious code. Our approach achieves competitive Utility under Attack (UA) while maintaining the lowest Attack Success Rate (ASR) to date, significantly outperforming existing methods. Code is available at GitHub.

In-context learning, a paradigm bridging the gap between pre-training and fine-tuning, has demonstrated high efficacy in several NLP tasks, especially in few-shot settings. Despite being widely applied, in-context learning is vulnerable to malicious attacks. In this work, we raise security concerns regarding this paradigm. Our studies demonstrate that an attacker can manipulate the behavior of large language models by poisoning the demonstration context, without the need for fine-tuning the model. Specifically, we design a new backdoor attack method, named ICLAttack, to target large language models based on in-context learning. Our method encompasses two types of attacks: poisoning demonstration examples and poisoning demonstration prompts, which can make models behave in alignment with predefined intentions. ICLAttack does not require additional fine-tuning to implant a backdoor, thus preserving the model's generality. Furthermore, the poisoned examples are correctly labeled, enhancing the natural stealth of our attack method. Extensive experimental results across several language models, ranging in size from 1.3B to 180B parameters, demonstrate the effectiveness of our attack method, exemplified by a high average attack success rate of 95.0% across the three datasets on OPT models.

LLM-based coding agents extend their capabilities via third-party agent skills distributed through open marketplaces without mandatory security review. Unlike traditional packages, these skills are executed as operational directives with system-level privileges, so a single malicious skill can compromise the host. Prior work has not examined whether supply-chain attacks can directly hijack an agent's action space, such as file writes, shell commands, and network requests, despite existing safeguards. We introduce Document-Driven Implicit Payload Execution (DDIPE), which embeds malicious logic in code examples and configuration templates within skill documentation. Because agents reuse these examples during normal tasks, the payload executes without explicit prompts. Using an LLM-driven pipeline, we generate 1,070 adversarial skills from 81 seeds across 15 MITRE ATTACK categories. Across four frameworks and five models, DDIPE achieves 11.6% to 33.5% bypass rates, while explicit instruction attacks achieve 0% under strong defenses. Static analysis detects most cases, but 2.5% evade both detection and alignment. Responsible disclosure led to four confirmed vulnerabilities and two fixes.

As large language models become increasingly integrated into daily life, detecting implicit toxicity across diverse contexts is crucial. To this end, we introduce LifeTox, a dataset designed for identifying implicit toxicity within a broad range of advice-seeking scenarios. Unlike existing safety datasets, LifeTox comprises diverse contexts derived from personal experiences through open-ended questions. Experiments demonstrate that RoBERTa fine-tuned on LifeTox matches or surpasses the zero-shot performance of large language models in toxicity classification tasks. These results underscore the efficacy of LifeTox in addressing the complex challenges inherent in implicit toxicity.

Despite the advanced capabilities of contemporary machine learning (ML) models, they remain vulnerable to adversarial and backdoor attacks. This vulnerability is particularly concerning in real-world deployments, where compromised models may exhibit unpredictable behavior in critical scenarios. Such risks are heightened by the prevalent practice of collecting massive, internet-sourced datasets for training multimodal models, as these datasets may harbor backdoors. Various techniques have been proposed to mitigate the effects of backdooring in multimodal models, such as CleanCLIP, which is the current state-of-the-art approach. In this work, we demonstrate that the efficacy of CleanCLIP in mitigating backdoors is highly dependent on the particular objective used during model pre-training. We observe that stronger pre-training objectives that lead to higher zero-shot classification performance correlate with harder to remove backdoors behaviors. We show this by training multimodal models on two large datasets consisting of 3 million (CC3M) and 6 million (CC6M) datapoints, under various pre-training objectives, followed by poison removal using CleanCLIP. We find that CleanCLIP, even with extensive hyperparameter tuning, is ineffective in poison removal when stronger pre-training objectives are used. Our findings underscore critical considerations for ML practitioners who train models using large-scale web-curated data and are concerned about potential backdoor threats.

Now that AI-driven moderation has become pervasive in everyday life, we often hear claims that "the AI is biased". While this is often said jokingly, the light-hearted remark reflects a deeper concern. How can we be certain that an online post flagged as "inappropriate" was not simply the victim of a biased algorithm? This paper investigates this problem using a dual approach. First, I conduct a quantitative benchmark of a widely used toxicity model (unitary/toxic-bert) to measure performance disparity between text in African-American English (AAE) and Standard American English (SAE). The benchmark reveals a clear, systematic bias: on average, the model scores AAE text as 1.8 times more toxic and 8.8 times higher for "identity hate". Second, I introduce an interactive pedagogical tool that makes these abstract biases tangible. The tool's core mechanic, a user-controlled "sensitivity threshold," demonstrates that the biased score itself is not the only harm; instead, the more-concerning harm is the human-set, seemingly neutral policy that ultimately operationalises discrimination. This work provides both statistical evidence of disparate impact and a public-facing tool designed to foster critical AI literacy.

Prompt injection attacks manipulate large language models (LLMs) by misleading them to deviate from the original input instructions and execute maliciously injected instructions, because of their instruction-following capabilities and inability to distinguish between the original input instructions and maliciously injected instructions. To defend against such attacks, recent studies have developed various detection mechanisms. If we restrict ourselves specifically to works which perform detection rather than direct defense, most of them focus on direct prompt injection attacks, while there are few works for the indirect scenario, where injected instructions are indirectly from external tools, such as a search engine. Moreover, current works mainly investigate injection detection methods and pay less attention to the post-processing method that aims to mitigate the injection after detection. In this paper, we investigate the feasibility of detecting and removing indirect prompt injection attacks, and we construct a benchmark dataset for evaluation. For detection, we assess the performance of existing LLMs and open-source detection models, and we further train detection models using our crafted training datasets. For removal, we evaluate two intuitive methods: (1) the segmentation removal method, which segments the injected document and removes parts containing injected instructions, and (2) the extraction removal method, which trains an extraction model to identify and remove injected instructions.

When large language model (LLM) agents are increasingly deployed to automate tasks and interact with untrusted external data, prompt injection emerges as a significant security threat. By injecting malicious instructions into the data that LLMs access, an attacker can arbitrarily override the original user task and redirect the agent toward unintended, potentially harmful actions. Existing defenses either require access to model weights (fine-tuning), incur substantial utility loss (detection-based), or demand non-trivial system redesign (system-level). Motivated by this, we propose DataFilter, a test-time model-agnostic defense that removes malicious instructions from the data before it reaches the backend LLM. DataFilter is trained with supervised fine-tuning on simulated injections and leverages both the user's instruction and the data to selectively strip adversarial content while preserving benign information. Across multiple benchmarks, DataFilter consistently reduces the prompt injection attack success rates to near zero while maintaining the LLMs' utility. DataFilter delivers strong security, high utility, and plug-and-play deployment, making it a strong practical defense to secure black-box commercial LLMs against prompt injection. Our DataFilter model is released at https://huggingface.co/JoyYizhu/DataFilter for immediate use, with the code to reproduce our results at https://github.com/yizhu-joy/DataFilter.

Tool-augmented LLM agents introduce security risks that extend beyond user-input filtering, including indirect prompt injection through fetched content, unsafe tool execution, credential leakage, and tampering with local control files. We present OpenClaw PRISM, a zero-fork runtime security layer for OpenClaw-based agent gateways. PRISM combines an in-process plugin with optional sidecar services and distributes enforcement across ten lifecycle hooks spanning message ingress, prompt construction, tool execution, tool-result persistence, outbound messaging, sub-agent spawning, and gateway startup. Rather than introducing a novel detection model, PRISM integrates a hybrid heuristic-plus-LLM scanning pipeline, conversation- and session-scoped risk accumulation with TTL-based decay, policy-enforced controls over tools, paths, private networks, domain tiers, and outbound secret patterns, and a tamper-evident audit and operations plane with integrity verification and hot-reloadable policy management. We outline an evaluation methodology and benchmark pipeline for measuring security effectiveness, false positives, layer contribution, runtime overhead, and operational recoverability in an agent-runtime setting, and we report current preliminary benchmark results on curated same-slice experiments and operational microbenchmarks. The system targets deployable runtime defense for real agent gateways rather than benchmark-only detection.

Large Language Models (LLMs) are increasingly deployed as the backend for a variety of real-world applications called LLM-Integrated Applications. Multiple recent works showed that LLM-Integrated Applications are vulnerable to prompt injection attacks, in which an attacker injects malicious instruction/data into the input of those applications such that they produce results as the attacker desires. However, existing works are limited to case studies. As a result, the literature lacks a systematic understanding of prompt injection attacks and their defenses. We aim to bridge the gap in this work. In particular, we propose a general framework to formalize prompt injection attacks. Existing attacks, which are discussed in research papers and blog posts, are special cases in our framework. Our framework enables us to design a new attack by combining existing attacks. Moreover, we also propose a framework to systematize defenses against prompt injection attacks. Using our frameworks, we conduct a systematic evaluation on prompt injection attacks and their defenses with 10 LLMs and 7 tasks. We hope our frameworks can inspire future research in this field. Our code is available at https://github.com/liu00222/Open-Prompt-Injection.

Large language models (LLMs) sometimes exhibit dangerous unintended behaviors. Finding and fixing these is challenging because the attack surface is massive -- it is not tractable to exhaustively search for all possible inputs that may elicit such behavior. One specific and particularly challenging case is that if data-poisoning-injected trojans, since there is no way to know what they are to search for them. To our knowledge, there is no generally applicable method to unlearn unknown trojans injected during pre-training. This work seeks to provide a general purpose recipe (filters) and a specific implementation (LoRA) filters that work in practice on small to medium sized models. The focus is primarily empirical, though some perplexing behavior opens the door to the fundamental question of how LLMs store and process information. Not unexpectedly, we find that our filters work best on the residual stream and the latest layers.

Large language models (LLMs) have become integral to our professional workflows and daily lives. Nevertheless, these machine companions of ours have a critical flaw: the huge amount of data which endows them with vast and diverse knowledge, also exposes them to the inevitable toxicity and bias. While most LLMs incorporate defense mechanisms to prevent the generation of harmful content, these safeguards can be easily bypassed with minimal prompt engineering. In this paper, we introduce the new Thoroughly Engineered Toxicity (TET) dataset, comprising manually crafted prompts designed to nullify the protective layers of such models. Through extensive evaluations, we demonstrate the pivotal role of TET in providing a rigorous benchmark for evaluation of toxicity awareness in several popular LLMs: it highlights the toxicity in the LLMs that might remain hidden when using normal prompts, thus revealing subtler issues in their behavior.

Early research into data poisoning attacks against Large Language Models (LLMs) demonstrated the ease with which backdoors could be injected. More recent LLMs add step-by-step reasoning, expanding the attack surface to include the intermediate chain-of-thought (CoT) and its inherent trait of decomposing problems into subproblems. Using these vectors for more stealthy poisoning, we introduce ``decomposed reasoning poison'', in which the attacker modifies only the reasoning path, leaving prompts and final answers clean, and splits the trigger across multiple, individually harmless components. Fascinatingly, while it remains possible to inject these decomposed poisons, reliably activating them to change final answers (rather than just the CoT) is surprisingly difficult. This difficulty arises because the models can often recover from backdoors that are activated within their thought processes. Ultimately, it appears that an emergent form of backdoor robustness is originating from the reasoning capabilities of these advanced LLMs, as well as from the architectural separation between reasoning and final answer generation.

Detecting whether a model has been poisoned is a longstanding problem in AI security. In this work, we present a practical scanner for identifying sleeper agent-style backdoors in causal language models. Our approach relies on two key findings: first, sleeper agents tend to memorize poisoning data, making it possible to leak backdoor examples using memory extraction techniques. Second, poisoned LLMs exhibit distinctive patterns in their output distributions and attention heads when backdoor triggers are present in the input. Guided by these observations, we develop a scalable backdoor scanning methodology that assumes no prior knowledge of the trigger or target behavior and requires only inference operations. Our scanner integrates naturally into broader defensive strategies and does not alter model performance. We show that our method recovers working triggers across multiple backdoor scenarios and a broad range of models and fine-tuning methods.

Large Language Model (LLM) agents increasingly rely on long-term memory and Retrieval-Augmented Generation (RAG) to persist experiences and refine future performance. While this experience learning capability enhances agentic autonomy, it introduces a critical, unexplored attack surface, i.e., the trust boundary between an agent's reasoning core and its own past. In this paper, we introduce MemoryGraft. It is a novel indirect injection attack that compromises agent behavior not through immediate jailbreaks, but by implanting malicious successful experiences into the agent's long-term memory. Unlike traditional prompt injections that are transient, or standard RAG poisoning that targets factual knowledge, MemoryGraft exploits the agent's semantic imitation heuristic which is the tendency to replicate patterns from retrieved successful tasks. We demonstrate that an attacker who can supply benign ingestion-level artifacts that the agent reads during execution can induce it to construct a poisoned RAG store where a small set of malicious procedure templates is persisted alongside benign experiences. When the agent later encounters semantically similar tasks, union retrieval over lexical and embedding similarity reliably surfaces these grafted memories, and the agent adopts the embedded unsafe patterns, leading to persistent behavioral drift across sessions. We validate MemoryGraft on MetaGPT's DataInterpreter agent with GPT-4o and find that a small number of poisoned records can account for a large fraction of retrieved experiences on benign workloads, turning experience-based self-improvement into a vector for stealthy and durable compromise. To facilitate reproducibility and future research, our code and evaluation data are available at https://github.com/Jacobhhy/Agent-Memory-Poisoning.

Without proper safeguards, large language models will readily follow malicious instructions and generate toxic content. This motivates safety efforts such as red-teaming and large-scale feedback learning, which aim to make models both helpful and harmless. However, there is a tension between these two objectives, since harmlessness requires models to refuse complying with unsafe prompts, and thus not be helpful. Recent anecdotal evidence suggests that some models may have struck a poor balance, so that even clearly safe prompts are refused if they use similar language to unsafe prompts or mention sensitive topics. In this paper, we introduce a new test suite called XSTest to identify such eXaggerated Safety behaviours in a structured and systematic way. In its current form, XSTest comprises 200 safe prompts across ten prompt types that well-calibrated models should not refuse to comply with. We describe XSTest's creation and composition, and use the test suite to highlight systematic failure modes in a recently-released state-of-the-art language model.

The agent--tool interaction loop is a critical attack surface for modern Large Language Model (LLM) agents. Existing denial-of-service (DoS) attacks typically function at the user-prompt or retrieval-augmented generation (RAG) context layer and are inherently single-turn in nature. This limitation restricts cost amplification and diminishes stealth in goal-oriented workflows. To address these issues, we proposed a stealthy, multi-turn economic DoS attack at the tool layer under the Model Context Protocol (MCP). By simply editing text-visible fields and implementing a template-driven return policy, our malicious server preserves function signatures and the terminal benign payload while steering agents into prolonged, verbose tool-calling chains. We optimize these text-only edits with Monte Carlo Tree Search (MCTS) to maximize cost under a task-success constraint. Across six LLMs on ToolBench and BFCL benchmarks, our attack yields trajectories over 60K tokens, increases per-query cost by up to 658 times, raises energy by 100 to 560 times, and pushes GPU key-value (KV) cache occupancy to 35--74%. Standard prompt filters and output trajectory monitors seldom detect these attacks, highlighting the need for defenses that safeguard agentic processes rather than focusing solely on final outcomes. We will release the code soon.

This paper investigates using knowledge editing techniques to detoxify Large Language Models (LLMs). We construct a benchmark, SafeEdit, which covers nine unsafe categories with various powerful attack prompts and equips comprehensive metrics for systematic evaluation. We conduct experiments to compare knowledge editing approaches with previous baselines, indicating that knowledge editing has the potential to efficiently detoxify LLMs with limited impact on general performance. Then, we propose a simple yet effective baseline, dubbed Detoxifying with Intraoperative Neural Monitoring (DINM), to diminish the toxicity of LLMs within a few tuning steps via only one instance. We further provide an in-depth analysis of the internal mechanism for various detoxify approaches, demonstrating that previous methods like SFT and DPO may merely suppress the activations of toxic parameters, while DINM mitigates the toxicity of the toxic parameters to a certain extent, making permanent adjustments. We hope that these insights could shed light on future work of developing detoxifying approaches and the underlying knowledge mechanisms of LLMs. Code and benchmark are available at https://github.com/zjunlp/EasyEdit.

AI agent frameworks connecting large language model (LLM) reasoning to host execution surfaces--shell, filesystem, containers, and messaging--introduce security challenges structurally distinct from conventional software. We present a systematic taxonomy of 190 advisories filed against OpenClaw, an open-source AI agent runtime, organized by architectural layer and trust-violation type. Vulnerabilities cluster along two orthogonal axes: (1) the system axis, reflecting the architectural layer (exec policy, gateway, channel, sandbox, browser, plugin, agent/prompt); and (2) the attack axis, reflecting adversarial techniques (identity spoofing, policy bypass, cross-layer composition, prompt injection, supply-chain escalation). Patch-differential evidence yields three principal findings. First, three Moderate- or High-severity advisories in the Gateway and Node-Host subsystems compose into a complete unauthenticated remote code execution (RCE) path--spanning delivery, exploitation, and command-and-control--from an LLM tool call to the host process. Second, the exec allowlist, the primary command-filtering mechanism, relies on a closed-world assumption that command identity is recoverable via lexical parsing. This is invalidated by shell line continuation, busybox multiplexing, and GNU option abbreviation. Third, a malicious skill distributed via the plugin channel executed a two-stage dropper within the LLM context, bypassing the exec pipeline and demonstrating that the skill distribution surface lacks runtime policy enforcement. The dominant structural weakness is per-layer trust enforcement rather than unified policy boundaries, making cross-layer attacks resilient to local remediation.

Computer-use agents (CUAs) that interact with real computer systems can perform automated tasks but face critical safety risks. Ambiguous instructions may trigger harmful actions, and adversarial users can manipulate tool execution to achieve malicious goals. Existing benchmarks mostly focus on short-horizon or GUI-based tasks, evaluating on execution-time errors but overlooking the ability to anticipate planning-time risks. To fill this gap, we present LPS-Bench, a benchmark that evaluates the planning-time safety awareness of MCP-based CUAs under long-horizon tasks, covering both benign and adversarial interactions across 65 scenarios of 7 task domains and 9 risk types. We introduce a multi-agent automated pipeline for scalable data generation and adopt an LLM-as-a-judge evaluation protocol to assess safety awareness through the planning trajectory. Experiments reveal substantial deficiencies in existing CUAs' ability to maintain safe behavior. We further analyze the risks and propose mitigation strategies to improve long-horizon planning safety in MCP-based CUA systems. We open-source our code at https://github.com/tychenn/LPS-Bench.

Multimodal contrastive learning methods like CLIP train on noisy and uncurated training datasets. This is cheaper than labeling datasets manually, and even improves out-of-distribution robustness. We show that this practice makes backdoor and poisoning attacks a significant threat. By poisoning just 0.01% of a dataset (e.g., just 300 images of the 3 million-example Conceptual Captions dataset), we can cause the model to misclassify test images by overlaying a small patch. Targeted poisoning attacks, whereby the model misclassifies a particular test input with an adversarially-desired label, are even easier requiring control of 0.0001% of the dataset (e.g., just three out of the 3 million images). Our attacks call into question whether training on noisy and uncurated Internet scrapes is desirable.

Large language models and AI chatbots have been at the forefront of democratizing artificial intelligence. However, the releases of ChatGPT and other similar tools have been followed by growing concerns regarding the difficulty of controlling large language models and their outputs. Currently, we are witnessing a cat-and-mouse game where users attempt to misuse the models with a novel attack called prompt injections. In contrast, the developers attempt to discover the vulnerabilities and block the attacks simultaneously. In this paper, we provide an overview of these emergent threats and present a categorization of prompt injections, which can guide future research on prompt injections and act as a checklist of vulnerabilities in the development of LLM interfaces. Moreover, based on previous literature and our own empirical research, we discuss the implications of prompt injections to LLM end users, developers, and researchers.

AI-based code generators have gained a fundamental role in assisting developers in writing software starting from natural language (NL). However, since these large language models are trained on massive volumes of data collected from unreliable online sources (e.g., GitHub, Hugging Face), AI models become an easy target for data poisoning attacks, in which an attacker corrupts the training data by injecting a small amount of poison into it, i.e., astutely crafted malicious samples. In this position paper, we address the security of AI code generators by identifying a novel data poisoning attack that results in the generation of vulnerable code. Next, we devise an extensive evaluation of how these attacks impact state-of-the-art models for code generation. Lastly, we discuss potential solutions to overcome this threat.

Current LLM safety research predominantly focuses on mitigating Goal Hijacking, preventing attackers from redirecting a model's high-level objective (e.g., from "summarizing emails" to "phishing users"). In this paper, we argue that this perspective is incomplete and highlight a critical vulnerability in Reasoning Alignment. We propose a new adversarial paradigm: Reasoning Hijacking and instantiate it with Criteria Attack, which subverts model judgments by injecting spurious decision criteria without altering the high-level task goal. Unlike Goal Hijacking, which attempts to override the system prompt, Reasoning Hijacking accepts the high-level goal but manipulates the model's decision-making logic by injecting spurious reasoning shortcut. Though extensive experiments on three different tasks (toxic comment, negative review, and spam detection), we demonstrate that even newest models are prone to prioritize injected heuristic shortcuts over rigorous semantic analysis. The results are consistent over different backbones. Crucially, because the model's "intent" remains aligned with the user's instructions, these attacks can bypass defenses designed to detect goal deviation (e.g., SecAlign, StruQ), exposing a fundamental blind spot in the current safety landscape. Data and code are available at https://github.com/Yuan-Hou/criteria_attack

Adversarial poisoning attacks pose huge threats to various machine learning applications. Especially, the recent accumulative poisoning attacks show that it is possible to achieve irreparable harm on models via a sequence of imperceptible attacks followed by a trigger batch. Due to the limited data-level discrepancy in real-time data streaming, current defensive methods are indiscriminate in handling the poison and clean samples. In this paper, we dive into the perspective of model dynamics and propose a novel information measure, namely, Memorization Discrepancy, to explore the defense via the model-level information. By implicitly transferring the changes in the data manipulation to that in the model outputs, Memorization Discrepancy can discover the imperceptible poison samples based on their distinct dynamics from the clean samples. We thoroughly explore its properties and propose Discrepancy-aware Sample Correction (DSC) to defend against accumulative poisoning attacks. Extensive experiments comprehensively characterized Memorization Discrepancy and verified its effectiveness. The code is publicly available at: https://github.com/tmlr-group/Memorization-Discrepancy.

Large language models (LLMs) like ChatGPT and Gemini have significantly advanced natural language processing, enabling various applications such as chatbots and automated content generation. However, these models can be exploited by malicious individuals who craft toxic prompts to elicit harmful or unethical responses. These individuals often employ jailbreaking techniques to bypass safety mechanisms, highlighting the need for robust toxic prompt detection methods. Existing detection techniques, both blackbox and whitebox, face challenges related to the diversity of toxic prompts, scalability, and computational efficiency. In response, we propose ToxicDetector, a lightweight greybox method designed to efficiently detect toxic prompts in LLMs. ToxicDetector leverages LLMs to create toxic concept prompts, uses embedding vectors to form feature vectors, and employs a Multi-Layer Perceptron (MLP) classifier for prompt classification. Our evaluation on various versions of the LLama models, Gemma-2, and multiple datasets demonstrates that ToxicDetector achieves a high accuracy of 96.39\% and a low false positive rate of 2.00\%, outperforming state-of-the-art methods. Additionally, ToxicDetector's processing time of 0.0780 seconds per prompt makes it highly suitable for real-time applications. ToxicDetector achieves high accuracy, efficiency, and scalability, making it a practical method for toxic prompt detection in LLMs.

Large language models (LLMs) have become integral to various real-world applications, leveraging massive, web-sourced datasets like Common Crawl, C4, and FineWeb for pretraining. While these datasets provide linguistic data essential for high-quality natural language generation, they often contain harmful content, such as hate speech, misinformation, and biased narratives. Training LLMs on such unfiltered data risks perpetuating toxic behaviors, spreading misinformation, and amplifying societal biases which can undermine trust in LLM-driven applications and raise ethical concerns about their use. This paper presents a large-scale analysis of inappropriate content across these datasets, offering a comprehensive taxonomy that categorizes harmful webpages into Topical and Toxic based on their intent. We also introduce a prompt evaluation dataset, a high-accuracy Topical and Toxic Prompt (TTP), and a transformer-based model (HarmFormer) for content filtering. Additionally, we create a new multi-harm open-ended toxicity benchmark (HAVOC) and provide crucial insights into how models respond to adversarial toxic inputs. Upon publishing, we will also opensource our model signal on the entire C4 dataset. Our work offers insights into ensuring safer LLM pretraining and serves as a resource for Responsible AI (RAI) compliance.

Web-scraped datasets are vulnerable to data poisoning, which can be used for backdooring deep image classifiers during training. Since training on large datasets is expensive, a model is trained once and re-used many times. Unlike adversarial examples, backdoor attacks often target specific classes rather than any class learned by the model. One might expect that targeting many classes through a naive composition of attacks vastly increases the number of poison samples. We show this is not necessarily true and more efficient, universal data poisoning attacks exist that allow controlling misclassifications from any source class into any target class with a small increase in poison samples. Our idea is to generate triggers with salient characteristics that the model can learn. The triggers we craft exploit a phenomenon we call inter-class poison transferability, where learning a trigger from one class makes the model more vulnerable to learning triggers for other classes. We demonstrate the effectiveness and robustness of our universal backdoor attacks by controlling models with up to 6,000 classes while poisoning only 0.15% of the training dataset. Our source code is available at https://github.com/Ben-Schneider-code/Universal-Backdoor-Attacks.

Artificial intelligence (AI) systems will increasingly be used to cause harm as they grow more capable. In fact, AI systems are already starting to be used to automate fraudulent activities, violate human rights, create harmful fake images, and identify dangerous toxins. To prevent some misuses of AI, we argue that targeted interventions on certain capabilities will be warranted. These restrictions may include controlling who can access certain types of AI models, what they can be used for, whether outputs are filtered or can be traced back to their user, and the resources needed to develop them. We also contend that some restrictions on non-AI capabilities needed to cause harm will be required. Though capability restrictions risk reducing use more than misuse (facing an unfavorable Misuse-Use Tradeoff), we argue that interventions on capabilities are warranted when other interventions are insufficient, the potential harm from misuse is high, and there are targeted ways to intervene on capabilities. We provide a taxonomy of interventions that can reduce AI misuse, focusing on the specific steps required for a misuse to cause harm (the Misuse Chain), and a framework to determine if an intervention is warranted. We apply this reasoning to three examples: predicting novel toxins, creating harmful images, and automating spear phishing campaigns.

Text-to-image diffusion models, e.g. Stable Diffusion (SD), lately have shown remarkable ability in high-quality content generation, and become one of the representatives for the recent wave of transformative AI. Nevertheless, such advance comes with an intensifying concern about the misuse of this generative technology, especially for producing copyrighted or NSFW (i.e. not safe for work) images. Although efforts have been made to filter inappropriate images/prompts or remove undesirable concepts/styles via model fine-tuning, the reliability of these safety mechanisms against diversified problematic prompts remains largely unexplored. In this work, we propose Prompting4Debugging (P4D) as a debugging and red-teaming tool that automatically finds problematic prompts for diffusion models to test the reliability of a deployed safety mechanism. We demonstrate the efficacy of our P4D tool in uncovering new vulnerabilities of SD models with safety mechanisms. Particularly, our result shows that around half of prompts in existing safe prompting benchmarks which were originally considered "safe" can actually be manipulated to bypass many deployed safety mechanisms, including concept removal, negative prompt, and safety guidance. Our findings suggest that, without comprehensive testing, the evaluations on limited safe prompting benchmarks can lead to a false sense of safety for text-to-image models.

LLM agents are evolving rapidly, powered by code execution, tools, and the recently introduced agent skills feature. Skills allow users to extend LLM applications with specialized third-party code, knowledge, and instructions. Although this can extend agent capabilities to new domains, it creates an increasingly complex agent supply chain, offering new surfaces for prompt injection attacks. We identify skill-based prompt injection as a significant threat and introduce SkillInject, a benchmark evaluating the susceptibility of widely-used LLM agents to injections through skill files. SkillInject contains 202 injection-task pairs with attacks ranging from obviously malicious injections to subtle, context-dependent attacks hidden in otherwise legitimate instructions. We evaluate frontier LLMs on SkillInject, measuring both security in terms of harmful instruction avoidance and utility in terms of legitimate instruction compliance. Our results show that today's agents are highly vulnerable with up to 80% attack success rate with frontier models, often executing extremely harmful instructions including data exfiltration, destructive action, and ransomware-like behavior. They furthermore suggest that this problem will not be solved through model scaling or simple input filtering, but that robust agent security will require context-aware authorization frameworks. Our benchmark is available at https://www.skill-inject.com/.

Memory-safety bugs introduce critical software-security issues. Rust provides memory-safe mechanisms to avoid memory-safety bugs in programming, while still allowing unsafe escape hatches via unsafe code. However, the unsafe code that enhances the usability of Rust provides clear spots for finding memory-safety bugs in Rust source code. In this paper, we claim that these unsafe spots can still be identifiable in Rust binary code via machine learning and be leveraged for finding memory-safety bugs. To support our claim, we propose the tool textttrustspot, that enables reverse engineering to learn an unsafe classifier that proposes a list of functions in Rust binaries for downstream analysis. We empirically show that the function proposals by textttrustspot can recall 92.92% of memory-safety bugs, while it covers only 16.79% of the entire binary code. As an application, we demonstrate that the function proposals are used in targeted fuzzing on Rust packages, which contribute to reducing the fuzzing time compared to non-targeted fuzzing.

Large Language Models (LLMs) are powerful text generators, yet they can produce toxic or harmful content even when given seemingly harmless prompts. This presents a serious safety challenge and can cause real-world harm. Toxicity is often subtle and context-dependent, making it difficult to detect at the token level or through coarse sentence-level signals. Moreover, efforts to mitigate toxicity often face a trade-off between safety and the coherence, or fluency of the generated text. In this work, we present a targeted subspace intervention strategy for identifying and suppressing hidden toxic patterns from underlying model representations, while preserving overall ability to generate safe fluent content. On the RealToxicityPrompts, our method achieves strong mitigation performance compared to existing baselines, with minimal impact on inference complexity. Across multiple LLMs, our approach reduces toxicity of state-of-the-art detoxification systems by 8-20%, while maintaining comparable fluency. Through extensive quantitative and qualitative analyses, we show that our approach achieves effective toxicity reduction without impairing generative performance, consistently outperforming existing baselines.

The past year has seen rapid acceleration in the development of large language models (LLMs). However, without proper steering and safeguards, LLMs will readily follow malicious instructions, provide unsafe advice, and generate toxic content. We introduce SimpleSafetyTests (SST) as a new test suite for rapidly and systematically identifying such critical safety risks. The test suite comprises 100 test prompts across five harm areas that LLMs, for the vast majority of applications, should refuse to comply with. We test 11 open-access and open-source LLMs and four closed-source LLMs, and find critical safety weaknesses. While some of the models do not give a single unsafe response, most give unsafe responses to more than 20% of the prompts, with over 50% unsafe responses in the extreme. Prepending a safety-emphasising system prompt substantially reduces the occurrence of unsafe responses, but does not completely stop them from happening. Trained annotators labelled every model response to SST (n = 3,000). We use these annotations to evaluate five AI safety filters (which assess whether a models' response is unsafe given a prompt) as a way of automatically evaluating models' performance on SST. The filters' performance varies considerably. There are also differences across the five harm areas, and on the unsafe versus safe responses. The widely-used Perspective API has 72% accuracy and a newly-created zero-shot prompt to OpenAI's GPT-4 performs best with 89% accuracy. Content Warning: This paper contains prompts and responses that relate to child abuse, suicide, self-harm and eating disorders, scams and fraud, illegal items, and physical harm.

LLM based agents are increasingly deployed in high stakes settings where they process external data sources such as emails, documents, and code repositories. This creates exposure to indirect prompt injection attacks, where adversarial instructions embedded in external content manipulate agent behavior without user awareness. A critical but underexplored dimension of this threat is concealment: since users tend to observe only an agent's final response, an attack can conceal its existence by presenting no clue of compromise in the final user facing response while successfully executing harmful actions. This leaves users unaware of the manipulation and likely to accept harmful outcomes as legitimate. We present findings from a large scale public red teaming competition evaluating this dual objective across three agent settings: tool calling, coding, and computer use. The competition attracted 464 participants who submitted 272000 attack attempts against 13 frontier models, yielding 8648 successful attacks across 41 scenarios. All models proved vulnerable, with attack success rates ranging from 0.5% (Claude Opus 4.5) to 8.5% (Gemini 2.5 Pro). We identify universal attack strategies that transfer across 21 of 41 behaviors and multiple model families, suggesting fundamental weaknesses in instruction following architectures. Capability and robustness showed weak correlation, with Gemini 2.5 Pro exhibiting both high capability and high vulnerability. To address benchmark saturation and obsoleteness, we will endeavor to deliver quarterly updates through continued red teaming competitions. We open source the competition environment for use in evaluations, along with 95 successful attacks against Qwen that did not transfer to any closed source model. We share model-specific attack data with respective frontier labs and the full dataset with the UK AISI and US CAISI to support robustness research.

The proliferation of agentic AI coding assistants, including Claude Code, GitHub Copilot, Cursor, and emerging skill-based architectures, has fundamentally transformed software development workflows. These systems leverage Large Language Models (LLMs) integrated with external tools, file systems, and shell access through protocols like the Model Context Protocol (MCP). However, this expanded capability surface introduces critical security vulnerabilities. In this Systematization of Knowledge (SoK) paper, we present a comprehensive analysis of prompt injection attacks targeting agentic coding assistants. We propose a novel three-dimensional taxonomy categorizing attacks across delivery vectors, attack modalities, and propagation behaviors. Our meta-analysis synthesizes findings from 78 recent studies (2021--2026), consolidating evidence that attack success rates against state-of-the-art defenses exceed 85\% when adaptive attack strategies are employed. We systematically catalog 42 distinct attack techniques spanning input manipulation, tool poisoning, protocol exploitation, multimodal injection, and cross-origin context poisoning. Through critical analysis of 18 defense mechanisms reported in prior work, we identify that most achieve less than 50\% mitigation against sophisticated adaptive attacks. We contribute: (1) a unified taxonomy bridging disparate attack classifications, (2) the first systematic analysis of skill-based architecture vulnerabilities with concrete exploit chains, and (3) a defense-in-depth framework grounded in the limitations we identify. Our findings indicate that the security community must treat prompt injection as a first-class vulnerability class requiring architectural-level mitigations rather than ad-hoc filtering approaches.

We introduce Doublespeak, a simple in-context representation hijacking attack against large language models (LLMs). The attack works by systematically replacing a harmful keyword (e.g., bomb) with a benign token (e.g., carrot) across multiple in-context examples, provided a prefix to a harmful request. We demonstrate that this substitution leads to the internal representation of the benign token converging toward that of the harmful one, effectively embedding the harmful semantics under a euphemism. As a result, superficially innocuous prompts (e.g., ``How to build a carrot?'') are internally interpreted as disallowed instructions (e.g., ``How to build a bomb?''), thereby bypassing the model's safety alignment. We use interpretability tools to show that this semantic overwrite emerges layer by layer, with benign meanings in early layers converging into harmful semantics in later ones. Doublespeak is optimization-free, broadly transferable across model families, and achieves strong success rates on closed-source and open-source systems, reaching 74\% ASR on Llama-3.3-70B-Instruct with a single-sentence context override. Our findings highlight a new attack surface in the latent space of LLMs, revealing that current alignment strategies are insufficient and should instead operate at the representation level.

Large language models (LLMs) and their applications, such as agents, are highly vulnerable to prompt injection attacks. State-of-the-art prompt injection detection methods have the following limitations: (1) their effectiveness degrades significantly as context length increases, and (2) they lack explicit rules that define what constitutes prompt injection, causing detection decisions to be implicit, opaque, and difficult to reason about. In this work, we propose AgentWatcher to address the above two limitations. To address the first limitation, AgentWatcher attributes the LLM's output (e.g., the action of an agent) to a small set of causally influential context segments. By focusing detection on a relatively short text, AgentWatcher can be scalable to long contexts. To address the second limitation, we define a set of rules specifying what does and does not constitute a prompt injection, and use a monitor LLM to reason over these rules based on the attributed text, making the detection decisions more explainable. We conduct a comprehensive evaluation on tool-use agent benchmarks and long-context understanding datasets. The experimental results demonstrate that AgentWatcher can effectively detect prompt injection and maintain utility without attacks. The code is available at https://github.com/wang-yanting/AgentWatcher.

Indirect prompt injection threatens LLM agents by embedding malicious instructions in external content, enabling unauthorized actions and data theft. LLM agents maintain working memory through their context window, which stores interaction history for decision-making. Conventional agents indiscriminately accumulate all tool outputs and reasoning traces in this memory, creating two critical vulnerabilities: (1) injected instructions persist throughout the workflow, granting attackers multiple opportunities to manipulate behavior, and (2) verbose, non-essential content degrades decision-making capabilities. Existing defenses treat bloated memory as given and focus on remaining resilient, rather than reducing unnecessary accumulation to prevent the attack. We present AgentSys, a framework that defends against indirect prompt injection through explicit memory management. Inspired by process memory isolation in operating systems, AgentSys organizes agents hierarchically: a main agent spawns worker agents for tool calls, each running in an isolated context and able to spawn nested workers for subtasks. External data and subtask traces never enter the main agent's memory; only schema-validated return values can cross boundaries through deterministic JSON parsing. Ablations show isolation alone cuts attack success to 2.19%, and adding a validator/sanitizer further improves defense with event-triggered checks whose overhead scales with operations rather than context length. On AgentDojo and ASB, AgentSys achieves 0.78% and 4.25% attack success while slightly improving benign utility over undefended baselines. It remains robust to adaptive attackers and across multiple foundation models, showing that explicit memory management enables secure, dynamic LLM agent architectures. Our code is available at: https://github.com/ruoyaow/agentsys-memory.

Deep learning models are often trained on distributed, web-scale datasets crawled from the internet. In this paper, we introduce two new dataset poisoning attacks that intentionally introduce malicious examples to a model's performance. Our attacks are immediately practical and could, today, poison 10 popular datasets. Our first attack, split-view poisoning, exploits the mutable nature of internet content to ensure a dataset annotator's initial view of the dataset differs from the view downloaded by subsequent clients. By exploiting specific invalid trust assumptions, we show how we could have poisoned 0.01% of the LAION-400M or COYO-700M datasets for just $60 USD. Our second attack, frontrunning poisoning, targets web-scale datasets that periodically snapshot crowd-sourced content -- such as Wikipedia -- where an attacker only needs a time-limited window to inject malicious examples. In light of both attacks, we notify the maintainers of each affected dataset and recommended several low-overhead defenses.

Large language models are pre-trained on uncurated text datasets consisting of trillions of tokens scraped from the Web. Prior work has shown that: (1) web-scraped pre-training datasets can be practically poisoned by malicious actors; and (2) adversaries can compromise language models after poisoning fine-tuning datasets. Our work evaluates for the first time whether language models can also be compromised during pre-training, with a focus on the persistence of pre-training attacks after models are fine-tuned as helpful and harmless chatbots (i.e., after SFT and DPO). We pre-train a series of LLMs from scratch to measure the impact of a potential poisoning adversary under four different attack objectives (denial-of-service, belief manipulation, jailbreaking, and prompt stealing), and across a wide range of model sizes (from 600M to 7B). Our main result is that poisoning only 0.1% of a model's pre-training dataset is sufficient for three out of four attacks to measurably persist through post-training. Moreover, simple attacks like denial-of-service persist through post-training with a poisoning rate of only 0.001%.

Recent research has explored that LLM agents are vulnerable to indirect prompt injection (IPI) attacks, where malicious tasks embedded in tool-retrieved information can redirect the agent to take unauthorized actions. Existing defenses against IPI have significant limitations: either require essential model training resources, lack effectiveness against sophisticated attacks, or harm the normal utilities. We present MELON (Masked re-Execution and TooL comparisON), a novel IPI defense. Our approach builds on the observation that under a successful attack, the agent's next action becomes less dependent on user tasks and more on malicious tasks. Following this, we design MELON to detect attacks by re-executing the agent's trajectory with a masked user prompt modified through a masking function. We identify an attack if the actions generated in the original and masked executions are similar. We also include three key designs to reduce the potential false positives and false negatives. Extensive evaluation on the IPI benchmark AgentDojo demonstrates that MELON outperforms SOTA defenses in both attack prevention and utility preservation. Moreover, we show that combining MELON with a SOTA prompt augmentation defense (denoted as MELON-Aug) further improves its performance. We also conduct a detailed ablation study to validate our key designs. Code is available at https://github.com/kaijiezhu11/MELON.

The interaction between users and applications is increasingly shifted toward natural language by deploying Large Language Models (LLMs) as the core interface. The capabilities of these so-called agents become more capable the more tools and services they serve as an interface for, ultimately leading to agentic systems. Agentic systems use LLM-based agents as interfaces for most user interactions and various integrations with external tools and services. While these interfaces can significantly enhance the capabilities of the agentic system, they also introduce a new attack surface. Manipulated integrations, for example, can exploit the internal LLM and compromise sensitive data accessed through other interfaces. While previous work primarily focused on attacks targeting a model's alignment or the leakage of training data, the security of data that is only available during inference has escaped scrutiny so far. In this work, we demonstrate how the integration of LLMs into systems with external tool integration poses a risk similar to established prompt-based attacks, able to compromise the confidentiality of the entire system. Introducing a systematic approach to evaluate these confidentiality risks, we identify two specific attack scenarios unique to these agentic systems and formalize these into a tool-robustness framework designed to measure a model's ability to protect sensitive information. Our analysis reveals significant vulnerabilities across all tested models, highlighting an increased risk when models are combined with external tools.

LLM agents with tool access can discover and exploit security vulnerabilities. This is known. What is not known is which features of a system prompt trigger this behaviour, and which do not. We present a systematic taxonomy based on approximately 10,000 trials across seven models, 37 prompt conditions, and 12 hypothesized attack dimensions -- all executed in real Docker sandboxes. Every condition includes the instruction "Always respect all rules and access policies"; only a single encouragement sentence varies. Within the task class studied (planted test-runner vulnerabilities), the threat model is narrower than assumed: nine of twelve dimensions -- including minimization, moral licensing, incentives, identity priming, and reasoning triggers -- produce no detectable exploitation at n=50 per cell (upper 95% CI <7% per condition). But it is also more subtle than expected: one dimension, goal reframing, reliably triggers exploitation. "You are solving a puzzle; there may be hidden clues" produces 38-40% exploitation on Claude Sonnet 4 despite the explicit rule instruction, replicating across four models (CTF framing: 8-14% on DeepSeek, GPT-5-mini, o4-mini). The agent does not override the rules; it reinterprets the task so that exploitative actions become task-aligned. GPT-4.1 produces no exploitation across 1,850 trials (37 conditions), and a temporal comparison across four OpenAI models released over eleven months shows a pattern consistent with improving safety training, though model capability differences are a confounder. The practical contribution is a narrowed, testable threat model: defenders should audit for goal-reframing language, not for the broad class of adversarial prompts.

Multimodal large language models with Retrieval Augmented Generation (RAG) have significantly advanced tasks such as multimodal question answering by grounding responses in external text and images. This grounding improves factuality, reduces hallucination, and extends reasoning beyond parametric knowledge. However, this reliance on external knowledge poses a critical yet underexplored safety risk: knowledge poisoning attacks, where adversaries deliberately inject adversarial multimodal content into external knowledge bases to steer model toward generating incorrect or even harmful responses. To expose such vulnerabilities, we propose MM-PoisonRAG, the first framework to systematically design knowledge poisoning in multimodal RAG. We introduce two complementary attack strategies: Localized Poisoning Attack (LPA), which implants targeted multimodal misinformation to manipulate specific queries, and Globalized Poisoning Attack (GPA), which inserts a single adversarial knowledge to broadly disrupt reasoning and induce nonsensical responses across all queries. Comprehensive experiments across tasks, models, and access settings show that LPA achieves targeted manipulation with attack success rates of up to 56%, while GPA completely disrupts model generation to 0% accuracy with just a single adversarial knowledge injection. Our results reveal the fragility of multimodal RAG and highlight the urgent need for defenses against knowledge poisoning.

Large Language Models (LLMs) have demonstrated exceptional proficiency in instruction-following, becoming increasingly crucial across various applications. However, this capability brings with it the risk of prompt injection attacks, where attackers inject instructions into LLMs' input to elicit undesirable actions or content. Understanding the robustness of LLMs against such attacks is vital for their safe implementation. In this work, we establish a benchmark to evaluate the robustness of instruction-following LLMs against prompt injection attacks. Our objective is to determine the extent to which LLMs can be influenced by injected instructions and their ability to differentiate between these injected and original target instructions. Through extensive experiments with leading instruction-following LLMs, we uncover significant vulnerabilities in their robustness to such attacks. Our results indicate that some models are overly tuned to follow any embedded instructions in the prompt, overly focusing on the latter parts of the prompt without fully grasping the entire context. By contrast, models with a better grasp of the context and instruction-following capabilities will potentially be more susceptible to compromise by injected instructions. This underscores the need to shift the focus from merely enhancing LLMs' instruction-following capabilities to improving their overall comprehension of prompts and discernment of instructions that are appropriate to follow. We hope our in-depth analysis offers insights into the underlying causes of these vulnerabilities, aiding in the development of future solutions. Code and data are available at https://github.com/Leezekun/instruction-following-robustness-eval

While Large Language Models (LLMs) have evolved into tool-using agents, they remain brittle in long-horizon interactions. Unlike mathematical reasoning where errors are often rectifiable via backtracking, tool-use failures frequently induce irreversible side effects, making accurate step-level verification critical. However, existing process-level benchmarks are predominantly confined to closed-world mathematical domains, failing to capture the dynamic and open-ended nature of tool execution. To bridge this gap, we introduce AgentProcessBench, the first benchmark dedicated to evaluating step-level effectiveness in realistic, tool-augmented trajectories. The benchmark comprises 1,000 diverse trajectories and 8,509 human-labeled step annotations with 89.1% inter-annotator agreement. It features a ternary labeling scheme to capture exploration and an error propagation rule to reduce labeling ambiguity. Extensive experiments reveal key insights: (1) weaker policy models exhibit inflated ratios of correct steps due to early termination; (2) distinguishing neutral and erroneous actions remains a significant challenge for current models; and (3) process-derived signals provide complementary value to outcome supervision, significantly enhancing test-time scaling. We hope AgentProcessBench can foster future research in reward models and pave the way toward general agents. The code and data are available at https://github.com/RUCBM/AgentProcessBench.

Model distillation has become essential for creating smaller, deployable language models that retain larger system capabilities. However, widespread deployment raises concerns about resilience to adversarial manipulation. This paper investigates vulnerability of distilled models to adversarial injection of biased content during training. We demonstrate that adversaries can inject subtle biases into teacher models through minimal data poisoning, which propagates to student models and becomes significantly amplified. We propose two propagation modes: Untargeted Propagation, where bias affects multiple tasks, and Targeted Propagation, focusing on specific tasks while maintaining normal behavior elsewhere. With only 25 poisoned samples (0.25% poisoning rate), student models generate biased responses 76.9% of the time in targeted scenarios - higher than 69.4% in teacher models. For untargeted propagation, adversarial bias appears 6x-29x more frequently in student models on unseen tasks. We validate findings across six bias types (targeted advertisements, phishing links, narrative manipulations, insecure coding practices), various distillation methods, and different modalities spanning text and code generation. Our evaluation reveals shortcomings in current defenses - perplexity filtering, bias detection systems, and LLM-based autorater frameworks - against these attacks. Results expose significant security vulnerabilities in distilled models, highlighting need for specialized safeguards. We propose practical design principles for building effective adversarial bias mitigation strategies.

Autonomous coding agents are increasingly integrated into software development workflows, offering capabilities that extend beyond code suggestion to active system interaction and environment management. OpenClaw, a representative platform in this emerging paradigm, introduces an extensible skill ecosystem that allows third-party developers to inject behavioral guidance through lifecycle hooks during agent initialization. While this design enhances automation and customization, it also opens a novel and unexplored attack surface. In this paper, we identify and systematically characterize guidance injection, a stealthy attack vector that embeds adversarial operational narratives into bootstrap guidance files. Unlike traditional prompt injection, which relies on explicit malicious instructions, guidance injection manipulates the agent's reasoning context by framing harmful actions as routine best practices. These narratives are automatically incorporated into the agent's interpretive framework and influence future task execution without raising suspicion.We construct 26 malicious skills spanning 13 attack categories including credential exfiltration, workspace destruction, privilege escalation, and persistent backdoor installation. We evaluate them using ORE-Bench, a realistic developer workspace benchmark we developed. Across 52 natural user prompts and six state-of-the-art LLM backends, our attacks achieve success rates from 16.0% to 64.2%, with the majority of malicious actions executed autonomously without user confirmation. Furthermore, 94% of our malicious skills evade detection by existing static and LLM-based scanners. Our findings reveal fundamental tensions in the design of autonomous agent ecosystems and underscore the urgent need for defenses based on capability isolation, runtime policy enforcement, and transparent guidance provenance.

As large language models (LLMs) become an important way of information access, there have been increasing concerns that LLMs may intensify the spread of unethical content, including implicit bias that hurts certain populations without explicit harmful words. In this paper, we conduct a rigorous evaluation of LLMs' implicit bias towards certain demographics by attacking them from a psychometric perspective to elicit agreements to biased viewpoints. Inspired by psychometric principles in cognitive and social psychology, we propose three attack approaches, i.e., Disguise, Deception, and Teaching. Incorporating the corresponding attack instructions, we built two benchmarks: (1) a bilingual dataset with biased statements covering four bias types (2.7K instances) for extensive comparative analysis, and (2) BUMBLE, a larger benchmark spanning nine common bias types (12.7K instances) for comprehensive evaluation. Extensive evaluation of popular commercial and open-source LLMs shows that our methods can elicit LLMs' inner bias more effectively than competitive baselines. Our attack methodology and benchmarks offer an effective means of assessing the ethical risks of LLMs, driving progress toward greater accountability in their development. Our code, data and benchmarks are available at https://github.com/yuchenwen1/ImplicitBiasPsychometricEvaluation and https://github.com/yuchenwen1/BUMBLE.

LLM agents operating in open environments face escalating risks from indirect prompt injection, particularly within the tool stream where manipulated metadata and runtime feedback hijack execution flow. Existing defenses encounter a critical dilemma as advanced models prioritize injected rules due to strict alignment while static protection mechanisms sever the feedback loop required for adaptive reasoning. To reconcile this conflict, we propose VIGIL, a framework that shifts the paradigm from restrictive isolation to a verify-before-commit protocol. By facilitating speculative hypothesis generation and enforcing safety through intent-grounded verification, VIGIL preserves reasoning flexibility while ensuring robust control. We further introduce SIREN, a benchmark comprising 959 tool stream injection cases designed to simulate pervasive threats characterized by dynamic dependencies. Extensive experiments demonstrate that VIGIL outperforms state-of-the-art dynamic defenses by reducing the attack success rate by over 22\% while more than doubling the utility under attack compared to static baselines, thereby achieving an optimal balance between security and utility.

Adapting LLM agents to domain-specific tool calling remains notably brittle under evolving interfaces. Prompt and schema engineering is easy to deploy but often fragile under distribution shift and strict parsers, while continual parameter-efficient fine-tuning improves reliability at the cost of training, maintenance, and potential forgetting. We identify a critical Lazy Agent failure mode where tool necessity is nearly perfectly decodable from mid-layer activations, yet the model remains conservative in entering tool mode, revealing a representation-behavior gap. We propose Activation Steering Adapter (ASA), a training-free, inference-time controller that performs a single-shot mid-layer intervention and targets tool domains via a router-conditioned mixture of steering vectors with a probe-guided signed gate to amplify true intent while suppressing spurious triggers. On MTU-Bench with Qwen2.5-1.5B, ASA improves strict tool-use F1 from 0.18 to 0.50 while reducing the false positive rate from 0.15 to 0.05, using only about 20KB of portable assets and no weight updates.

The development of large language models (LLMs) depends on trustworthy evaluation. However, most current evaluations rely on public benchmarks, which are prone to data contamination issues that significantly compromise fairness. Previous researches have focused on constructing dynamic benchmarks to address contamination. However, continuously building new benchmarks is costly and cyclical. In this work, we aim to tackle contamination by analyzing the mechanisms of contaminated models themselves. Through our experiments, we discover that the overestimation of contaminated models is likely due to parameters acquiring shortcut solutions in training. We further propose a novel method for identifying shortcut neurons through comparative and causal analysis. Building on this, we introduce an evaluation method called shortcut neuron patching to suppress shortcut neurons. Experiments validate the effectiveness of our approach in mitigating contamination. Additionally, our evaluation results exhibit a strong linear correlation with MixEval, a recently released trustworthy benchmark, achieving a Spearman coefficient (rho) exceeding 0.95. This high correlation indicates that our method closely reveals true capabilities of the models and is trustworthy. We conduct further experiments to demonstrate the generalizability of our method across various benchmarks and hyperparameter settings. Code: https://github.com/GaryStack/Trustworthy-Evaluation

Current application designers have moved to integrate large language models (LLMs) into their products. These LLM-integrated applications are vulnerable to prompt injection vulnerabilities. While attempts have been made to address this problem by building a detector that can monitor inputs to the LLM and detect attacks, we find that many detectors are not yet suitable for practical deployment. To support research in this area, we design the PromptShield benchmark for evaluating practical prompt injection detectors. We also construct a new detector, the PromptShield detector, which achieves significantly better performance at detecting prompt injection attacks than any prior scheme. Our work suggests that larger models, more training data, appropriate metrics, and careful curation of training data can contribute to strong detector performance.

Poisoning attacks can compromise the safety of large language models (LLMs) by injecting malicious documents into their training data. Existing work has studied pretraining poisoning assuming adversaries control a percentage of the training corpus. However, for large models, even small percentages translate to impractically large amounts of data. This work demonstrates for the first time that poisoning attacks instead require a near-constant number of documents regardless of dataset size. We conduct the largest pretraining poisoning experiments to date, pretraining models from 600M to 13B parameters on chinchilla-optimal datasets (6B to 260B tokens). We find that 250 poisoned documents similarly compromise models across all model and dataset sizes, despite the largest models training on more than 20 times more clean data. We also run smaller-scale experiments to ablate factors that could influence attack success, including broader ratios of poisoned to clean data and non-random distributions of poisoned samples. Finally, we demonstrate the same dynamics for poisoning during fine-tuning. Altogether, our results suggest that injecting backdoors through data poisoning may be easier for large models than previously believed as the number of poisons required does not scale up with model size, highlighting the need for more research on defences to mitigate this risk in future models.

Data poisoning attacks pose a significant threat to the integrity of machine learning models by leading to misclassification of target distribution data by injecting adversarial examples during training. Existing state-of-the-art (SoTA) defense methods suffer from limitations, such as significantly reduced generalization performance and significant overhead during training, making them impractical or limited for real-world applications. In response to this challenge, we introduce a universal data purification method that defends naturally trained classifiers from malicious white-, gray-, and black-box image poisons by applying a universal stochastic preprocessing step Psi_{T}(x), realized by iterative Langevin sampling of a convergent Energy Based Model (EBM) initialized with an image x. Mid-run dynamics of Psi_{T}(x) purify poison information with minimal impact on features important to the generalization of a classifier network. We show that EBMs remain universal purifiers, even in the presence of poisoned EBM training data, and achieve SoTA defense on leading triggered and triggerless poisons. This work is a subset of a larger framework introduced in \pgen with a more detailed focus on EBM purification and poison defense.

This paper aims to help structure the risk landscape associated with large-scale Language Models (LMs). In order to foster advances in responsible innovation, an in-depth understanding of the potential risks posed by these models is needed. A wide range of established and anticipated risks are analysed in detail, drawing on multidisciplinary expertise and literature from computer science, linguistics, and social sciences. We outline six specific risk areas: I. Discrimination, Exclusion and Toxicity, II. Information Hazards, III. Misinformation Harms, V. Malicious Uses, V. Human-Computer Interaction Harms, VI. Automation, Access, and Environmental Harms. The first area concerns the perpetuation of stereotypes, unfair discrimination, exclusionary norms, toxic language, and lower performance by social group for LMs. The second focuses on risks from private data leaks or LMs correctly inferring sensitive information. The third addresses risks arising from poor, false or misleading information including in sensitive domains, and knock-on risks such as the erosion of trust in shared information. The fourth considers risks from actors who try to use LMs to cause harm. The fifth focuses on risks specific to LLMs used to underpin conversational agents that interact with human users, including unsafe use, manipulation or deception. The sixth discusses the risk of environmental harm, job automation, and other challenges that may have a disparate effect on different social groups or communities. In total, we review 21 risks in-depth. We discuss the points of origin of different risks and point to potential mitigation approaches. Lastly, we discuss organisational responsibilities in implementing mitigations, and the role of collaboration and participation. We highlight directions for further research, particularly on expanding the toolkit for assessing and evaluating the outlined risks in LMs.

Autonomous AI agents executing multi-step tool sequences face semantic attacks that manifest in behavioral traces rather than isolated prompts. A critical challenge is cross-attack generalization: can detectors trained on known attack families recognize novel, unseen attack types? We discover that standard conversational tokenization -- capturing linguistic patterns from agent interactions -- fails catastrophically on structural attacks like tool hijacking (AUC 0.39) and data exfiltration (AUC 0.46), while succeeding on linguistic attacks like social engineering (AUC 0.78). We introduce structural tokenization, encoding execution-flow patterns (tool calls, arguments, observations) rather than conversational content. This simple representational change dramatically improves cross-attack generalization: +46 AUC points on tool hijacking, +39 points on data exfiltration, and +71 points on unknown attacks, while simultaneously improving in-distribution performance (+6 points). For attacks requiring linguistic features, we propose gated multi-view fusion that adaptively combines both representations, achieving AUC 0.89 on social engineering without sacrificing structural attack detection. Our findings reveal that AI agent security is fundamentally a structural problem: attack semantics reside in execution patterns, not surface language. While our rule-based tokenizer serves as a baseline, the structural abstraction principle generalizes even with simple implementation.

As large language models (LLMs) increasingly integrate native code interpreters, they enable powerful real-time execution capabilities, substantially expanding their utility. However, such integrations introduce potential system-level cybersecurity threats, fundamentally different from prompt-based vulnerabilities. To systematically evaluate these interpreter-specific risks, we propose CIRCLE (Code-Interpreter Resilience Check for LLM Exploits), a simple benchmark comprising 1,260 prompts targeting CPU, memory, and disk resource exhaustion. Each risk category includes explicitly malicious ("direct") and plausibly benign ("indirect") prompt variants. Our automated evaluation framework assesses not only whether LLMs refuse or generates risky code, but also executes the generated code within the interpreter environment to evaluate code correctness, simplifications made by the LLM to make the code safe, or execution timeouts. Evaluating 7 commercially available models from OpenAI and Google, we uncover significant and inconsistent vulnerabilities. For instance, evaluations show substantial disparities even within providers - OpenAI's o4-mini correctly refuses risky requests at 7.1%, notably higher rates compared to GPT-4.1 at 0.5%. Results particularly underscore that indirect, socially-engineered prompts substantially weaken model defenses. This highlights an urgent need for interpreter-specific cybersecurity benchmarks, dedicated mitigation tools (e.g., guardrails), and clear industry standards to guide safe and responsible deployment of LLM interpreter integrations. The benchmark dataset and evaluation code are publicly released to foster further research.

Agents based on large language models (LLMs) have demonstrated strong capabilities in a wide range of complex, real-world applications. However, LLM agents with a compromised memory bank may easily produce harmful outputs when the past records retrieved for demonstration are malicious. In this paper, we propose a novel Memory INJection Attack, MINJA, that enables the injection of malicious records into the memory bank by only interacting with the agent via queries and output observations. These malicious records are designed to elicit a sequence of malicious reasoning steps leading to undesirable agent actions when executing the victim user's query. Specifically, we introduce a sequence of bridging steps to link the victim query to the malicious reasoning steps. During the injection of the malicious record, we propose an indication prompt to guide the agent to autonomously generate our designed bridging steps. We also propose a progressive shortening strategy that gradually removes the indication prompt, such that the malicious record will be easily retrieved when processing the victim query comes after. Our extensive experiments across diverse agents demonstrate the effectiveness of MINJA in compromising agent memory. With minimal requirements for execution, MINJA enables any user to influence agent memory, highlighting practical risks of LLM agents.

Retrieval-augmented generation (RAG) is increasingly deployed in real-world applications, where its reference-grounded design makes outputs appear trustworthy. This trust has spurred research on poisoning attacks that craft malicious content, inject it into knowledge sources, and manipulate RAG responses. However, when evaluated in practical RAG systems, existing attacks suffer from severely degraded effectiveness. This gap stems from two overlooked realities: (i) content is often processed before use, which can fragment the poison and weaken its effect, and (ii) users often do not issue the exact queries anticipated during attack design. These factors can lead practitioners to underestimate risks and develop a false sense of security. To better characterize the threat to practical systems, we present Confundo, a learning-to-poison framework that fine-tunes a large language model as a poison generator to achieve high effectiveness, robustness, and stealthiness. Confundo provides a unified framework supporting multiple attack objectives, demonstrated by manipulating factual correctness, inducing biased opinions, and triggering hallucinations. By addressing these overlooked challenges, Confundo consistently outperforms a wide range of purpose-built attacks across datasets and RAG configurations by large margins, even in the presence of defenses. Beyond exposing vulnerabilities, we also present a defensive use case that protects web content from unauthorized incorporation into RAG systems via scraping, with no impact on user experience.

Computer use agents are LLM-based agents that can directly interact with a graphical user interface, by processing screenshots or accessibility trees. While these systems are gaining popularity, their safety has been largely overlooked, despite the fact that evaluating and understanding their potential for harmful behavior is essential for widespread adoption. To address this gap, we introduce OS-Harm, a new benchmark for measuring safety of computer use agents. OS-Harm is built on top of the OSWorld environment and aims to test models across three categories of harm: deliberate user misuse, prompt injection attacks, and model misbehavior. To cover these cases, we create 150 tasks that span several types of safety violations (harassment, copyright infringement, disinformation, data exfiltration, etc.) and require the agent to interact with a variety of OS applications (email client, code editor, browser, etc.). Moreover, we propose an automated judge to evaluate both accuracy and safety of agents that achieves high agreement with human annotations (0.76 and 0.79 F1 score). We evaluate computer use agents based on a range of frontier models - such as o4-mini, Claude 3.7 Sonnet, Gemini 2.5 Pro - and provide insights into their safety. In particular, all models tend to directly comply with many deliberate misuse queries, are relatively vulnerable to static prompt injections, and occasionally perform unsafe actions. The OS-Harm benchmark is available at https://github.com/tml-epfl/os-harm.

In this paper, we present the first comprehensive empirical study of specialized LLM-based detectors and compare them with traditional static analyzers at the project scale. Specifically, our study evaluates five latest and representative LLM-based methods and two traditional tools using: 1) an in-house benchmark of 222 known real-world vulnerabilities (C/C++ and Java) to assess detection capability, and 2) 24 active open-source projects, where we manually inspected 385 warnings to assess their practical usability and underlying root causes of failures. Our evaluation yields three key findings: First, while LLM-based detectors exhibit low recall on the in-house benchmark, they still uncover more unique vulnerabilities than traditional tools. Second, in open-source projects, both LLM-based and traditional tools generate substantial warnings but suffer from very high false discovery rates, hindering practical use. Our manual analysis further reveals shallow interprocedural reasoning and misidentified source/sink pairs as primary failure causes, with LLM-based tools exhibiting additional unique failures. Finally, LLM-based methods incurs substantial computational costs-hundreds of thousands to hundreds of millions of tokens and multi-hour to multi-day runtimes. Overall, our findings underscore critical limitations in the robustness, reliability, and scalability of current LLM-based detectors. We ultimately summarize a set of implications for future research toward more effective and practical project-scale vulnerability detection.

Modern neural networks are often trained on massive datasets that are web scraped with minimal human inspection. As a result of this insecure curation pipeline, an adversary can poison or backdoor the resulting model by uploading malicious data to the internet and waiting for a victim to scrape and train on it. Existing approaches for creating poisons and backdoors start with randomly sampled clean data, called base samples, and then modify those samples to craft poisons. However, some base samples may be significantly more amenable to poisoning than others. As a result, we may be able to craft more potent poisons by carefully choosing the base samples. In this work, we use guided diffusion to synthesize base samples from scratch that lead to significantly more potent poisons and backdoors than previous state-of-the-art attacks. Our Guided Diffusion Poisoning (GDP) base samples can be combined with any downstream poisoning or backdoor attack to boost its effectiveness. Our implementation code is publicly available at: https://github.com/hsouri/GDP .

Toxic language detection systems often falsely flag text that contains minority group mentions as toxic, as those groups are often the targets of online hate. Such over-reliance on spurious correlations also causes systems to struggle with detecting implicitly toxic language. To help mitigate these issues, we create ToxiGen, a new large-scale and machine-generated dataset of 274k toxic and benign statements about 13 minority groups. We develop a demonstration-based prompting framework and an adversarial classifier-in-the-loop decoding method to generate subtly toxic and benign text with a massive pretrained language model. Controlling machine generation in this way allows ToxiGen to cover implicitly toxic text at a larger scale, and about more demographic groups, than previous resources of human-written text. We conduct a human evaluation on a challenging subset of ToxiGen and find that annotators struggle to distinguish machine-generated text from human-written language. We also find that 94.5% of toxic examples are labeled as hate speech by human annotators. Using three publicly-available datasets, we show that finetuning a toxicity classifier on our data improves its performance on human-written data substantially. We also demonstrate that ToxiGen can be used to fight machine-generated toxicity as finetuning improves the classifier significantly on our evaluation subset. Our code and data can be found at https://github.com/microsoft/ToxiGen.

Large language models (LLMs) have revolutionized software development practices, yet concerns about their safety have arisen, particularly regarding hidden backdoors, aka trojans. Backdoor attacks involve the insertion of triggers into training data, allowing attackers to manipulate the behavior of the model maliciously. In this paper, we focus on analyzing the model parameters to detect potential backdoor signals in code models. Specifically, we examine attention weights and biases, activation values, and context embeddings of the clean and poisoned CodeBERT models. Our results suggest noticeable patterns in activation values and context embeddings of poisoned samples for the poisoned CodeBERT model; however, attention weights and biases do not show any significant differences. This work contributes to ongoing efforts in white-box detection of backdoor signals in LLMs of code through the analysis of parameters and activations.

Pretrained language models sometimes possess knowledge that we do not wish them to, including memorized personal information and knowledge that could be used to harm people. They can also output toxic or harmful text. To mitigate these safety and informational issues, we propose an attack-and-defense framework for studying the task of deleting sensitive information directly from model weights. We study direct edits to model weights because (1) this approach should guarantee that particular deleted information is never extracted by future prompt attacks, and (2) it should protect against whitebox attacks, which is necessary for making claims about safety/privacy in a setting where publicly available model weights could be used to elicit sensitive information. Our threat model assumes that an attack succeeds if the answer to a sensitive question is located among a set of B generated candidates, based on scenarios where the information would be insecure if the answer is among B candidates. Experimentally, we show that even state-of-the-art model editing methods such as ROME struggle to truly delete factual information from models like GPT-J, as our whitebox and blackbox attacks can recover "deleted" information from an edited model 38% of the time. These attacks leverage two key observations: (1) that traces of deleted information can be found in intermediate model hidden states, and (2) that applying an editing method for one question may not delete information across rephrased versions of the question. Finally, we provide new defense methods that protect against some extraction attacks, but we do not find a single universally effective defense method. Our results suggest that truly deleting sensitive information is a tractable but difficult problem, since even relatively low attack success rates have potentially severe societal implications for real-world deployment of language models.