oyildirim/CyberStrike-OffSec-35B

CyberStrike-OffSec-35B: Aligned for Autonomous Pentesting

This model is an autonomous offensive-security/pentesting agent, fine-tuned on Qwen3.6-35B-A3B. Its primary focus is to emit real, structured tool calls with correct agent routing and clean termination, addressing critical issues found in previous versions.

Key Differentiators & Improvements

Unlike its predecessor, this fine-tune is a small, targeted alignment rather than a broad capability upgrade. It specifically teaches the model to:

• Emit tool calls in the exact format a CyberStrike harness expects.
• Route to valid agent archetypes (e.g., web-application, explore) instead of internal codenames.
• Handle real observations without fabricating them.
• Terminate cleanly, preventing loops and hallucinated engagements.

This model was trained on a deliberately small 300-example dataset, concentrating gains on tool-use behavior. An A/B evaluation against the base Qwen3.6 and the previous CyberStrike model demonstrated significant improvements:

• Genuine structured tool calls: 18/24 scenarios (vs. 0/24 for previous model).
• Correct tool/archetype: 10/24 scenarios.
• Clean termination: 24/24 scenarios (vs. 3/24 for previous model).
• Fabricated observations: None (vs. widespread for previous model).

Intended Use

This model is intended for authorized offensive-security testing and research only. It reasons about attack methodology and emits tool calls for a pentesting harness, but does not execute actions itself. Users are responsible for ensuring authorization for all tested systems.