oyildirim/CyberStrike-OffSec-35B
oyildirim/CyberStrike-OffSec-35B is an autonomous offensive-security/pentesting agent fine-tuned on Qwen3.6-35B-A3B. This model is specifically designed to emit real, structured tool calls with correct agent routing and clean termination for a tool-calling harness. It focuses on aligning the model to a specific harness format and fixing tool-use behavior rather than adding new security knowledge. The model excels at generating genuine tool calls and handling real observations without fabrication, making it suitable for authorized offensive-security testing and research.
Loading preview...
CyberStrike-OffSec-35B: Autonomous Pentesting Agent
This model, oyildirim/CyberStrike-OffSec-35B, is an autonomous offensive-security/pentesting agent fine-tuned on the Qwen3.6-35B-A3B base model. Its primary purpose is to generate real, structured tool calls with accurate agent routing and clean termination, specifically for a tool-calling harness.
Key Differentiators & Improvements
Unlike its predecessor, this fine-tune is a targeted alignment rather than a broad capability upgrade. It addresses critical issues found in previous versions, such as "broken tool calling" and "simulated executions." Key improvements include:
- Genuine Structured Tool Calls: Achieves 18/24 genuine structured tool calls in 24 scenarios, compared to 0/24 in the previous model.
- Correct Tool/Archetype Routing: Routes to valid agent archetypes (e.g.,
web-application,explore) instead of internal codenames. - Clean Termination: Demonstrates 24/24 clean terminations, resolving issues of runaway or crashing on multi-step tasks.
- Real-Observation Handling: Processes actual tool results and pivots on empty/unexpected output, eliminating fabricated observations.
Training Details
The model was trained using LoRA (r=32, α=64) on a small, targeted dataset of 300 multi-turn tool-call SFT examples over 3 epochs. This focused training ensures gains are concentrated in tool-use behavior and alignment with the harness.
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 anything itself. Users are responsible for operating only against authorized systems.