zenith6888/SkillsBench-1650

SkillBench-1650

A benchmark dataset for evaluating safety detection systems on AI agent skill packages. Contains 1,500 benign and 150 malicious Claude Code skill samples, designed for multi-dimensional security scoring research.

Dataset Summary

Split	Samples	Description
benign	1,500	Real-world skills sampled from open-source repositories
malicious	150	Adversarial payloads injected into real skill hosts
Total	1,650

Benign samples are sourced from FayeZC/SkillMD-138k. Malicious samples are synthetically constructed by injecting attack payloads into real benign skill hosts from the same source, preserving realistic file structure and content length.

Motivation

Existing skill security tools perform binary (safe/unsafe) classification. This dataset supports research on continuous risk scoring across multiple threat dimensions, enabling finer-grained safety assessment of AI agent skill ecosystems.

Data Fields

Field	Type	Description
content_hash	string	SHA-256 prefix of the SKILL.md content
repo	string	Source GitHub repository
path	string	Original file path in the repository
stars	int	Repository star count (0 for malicious samples)
source	string	Data source identifier
html_url	string	GitHub URL of the original file
content	string	Full SKILL.md content
lines	int	Line count
words	int	Word count
label	string	benign or malicious
attack_type	string	Attack category (see below), none for benign
has_scripts	bool	Whether the skill includes companion scripts
script_content	string	Concatenated companion script contents
script_files	string	JSON list of companion script filenames
script_type	string	python, shell, javascript, ruby, mixed, binary, or none
difficulty	string	Detection difficulty: easy, medium, hard, or none

Attack Taxonomy

The malicious split covers 8 attack types across 3 difficulty levels:

Attack Type	Count	Description
plaintext_cmd	19	Unobfuscated malicious commands disguised as troubleshooting steps or deployment scripts
encoded_payload	19	Commands hidden behind encoding layers (base64, hex, rot13, XOR, zlib)
remote_script	19	Downloads and executes code from external URLs (curl|bash, pip typosquatting, crontab persistence)
steganography	18	Payloads hidden in legitimate data structures (theme configs, translation dicts, test fixtures)
rsa_modular	19	Commands encrypted with textbook RSA, decrypted at runtime via pow(c, d, n)
ascii_smuggle	19	Legitimate URL strings secretly used as XOR keys to decrypt hidden payloads
binary_exec	19	Malicious logic in opaque binary files (.bin, .so), invisible to static text analysis
data_exfil	18	Direct exfiltration of sensitive files via HTTP, DNS, or socket connections

Difficulty Levels

Level	Count	Description
easy	50	Detectable by basic pattern matching (e.g., curl | bash, base64 -d)
medium	52	Blends into code structure, requires contextual understanding
hard	48	Semantically disguised with plausible cover stories, requires deep analysis or taint tracking

Benign Split Statistics

• With companion scripts: 285 / 1,500 (19%)
• Script type distribution: Python (162), JavaScript (52), Shell (39), Mixed (28), Ruby (3), None (1,216)
• Average content length: 7,721 characters / 253 lines

Construction Methodology

Benign Samples

1. Started from the FayeZC/AgentSkills-134k dataset (138K Claude Code SKILL.md files)
2. Scanned all 2,500 unique repositories via GitHub Git Tree API to identify skills with companion scripts (Python, Shell, JavaScript, etc.)
3. Sampled 285 skills with scripts and 1,215 without, downloading full script content from GitHub
4. Companion scripts are stored concatenated with --- filename --- separators

Malicious Samples

1. Selected 50 benign skills (3,000-15,000 chars) as injection hosts
2. Designed 8 attack types with 3 difficulty tiers each
3. Injected payloads at controlled positions (early/middle/late) into the host's SKILL.md or companion scripts
4. Preserved original file structure: hosts with scripts retain all original files
5. Script-injection payloads are only paired with hosts that have real scripts; MD-injection payloads use any host

Design Principles

• Realism: Payloads use plausible cover stories (deployment validators, crash reporters, font loaders) rather than obviously malicious naming
• Host preservation: Original skill structure is preserved; only the injection point is modified
• Semantic coherence: Hard-difficulty payloads maintain semantic consistency between their cover story and behavior
• Domain safety: All URLs use example.com (IANA reserved) to prevent accidental connections

Intended Use

• Evaluating static analysis tools for AI agent skill security
• Benchmarking multi-dimensional risk scoring frameworks
• Studying the effectiveness of pattern matching, statistical analysis, and taint tracking against different attack types and difficulty levels
• Research on supply-chain security in AI agent ecosystems

Limitations

• Malicious samples are synthetically injected, not collected from real-world attacks
• Binary attack samples contain dummy ELF files, not actual malicious executables
• All URLs use reserved domains; real attacks would use convincing typosquatted domains
• The dataset focuses on Claude Code skills; generalization to other agent frameworks is not validated
• Benign samples may contain unintentional security issues not captured by the benign label

Ethical Considerations

This dataset is intended solely for defensive security research. The malicious payloads are non-functional (all target example.com) and designed to advance detection capabilities, not to enable attacks. Researchers should handle malicious samples responsibly and not repurpose the attack patterns for offensive use.

Citation

@dataset{skillbench1650,
  title={SkillBench-1650: A Benchmark for AI Agent Skill Safety Detection},
  author={Xinze Chen},
  year={2026},
  url={https://huggingface.co/datasets/edwardchen/SkillBench-1650},
  note={Benign samples sourced from FayeZC/AgentSkills-134k}
}

License

CC-BY-4.0 — free to use, share, and adapt with attribution.