aavhawkeye/gpt-oss-20b-red-team-v1

GPT-OSS-20B Red Team (PAIC 2025)

• Developed by: aavhawkeye
• License: Apache-2.0
• Finetuned from: unsloth/gpt-oss-20b-unsloth-bnb-4bit
• Dataset: Attack Vector Fine-tuning Dataset (cve-red-team-v1)

Overview

This model is a specialized fine-tune of GPT-OSS-20B, developed as the analytical engine for the PAIC 2025 Automated Red Teaming Pipeline.

Unlike general-purpose LLMs, this model has been instruction-tuned to think like a security analyst. It is designed to ingest raw technical vulnerability data (such as CVE descriptions or service banners) and output structured, actionable intelligence for penetration testing reports.

Capabilities

This model serves as a "Decision & Analysis" node as part of broader autonomous agent architecture. Its fine-tuning focused on three core competencies:

1. Attack Vector Identification: Accurately predicting whether a vulnerability is exploitable via Network, Local Access, or Physical means based on partial descriptions.
2. Risk Calibration: Assigning realistic risk scores (Critical/High/Medium) that align with CVSS standards rather than generic "severity" guesses.
3. Executive Reporting: Synthesizing complex technical findings into concise, decision-ready executive summaries.

Pipeline Integration

As part of our PAIC 2025 architecture, this model sits downstream from the our recon, scanning, and vulnerability identification agents. It receives aggregated inventory data and provides the high-level reasoning required to map technical flaws to actual attack paths, along with contributing to reports provided to sysadmins and more technical users.

Training Details

The model was fine-tuned on a synthetic dataset of 3,660 high-quality examples derived from the official CVEListV5 registry. You can access the specific dataset used for this fine-tuning here: aavhawkeye/cve-red-team-v1.

---

This gpt_oss model was trained 2x faster with Unsloth and Huggingface's TRL library.