huihui-ai Strips Refusals From CyberStrike-OffSec-35B for Local Pentest Agents
An uncensored offensive-security model fine-tuned on Qwen3.6-35B lands on Hugging Face, aimed at autonomous pentesting agents with tool calling.
Takeaways
- Huihui-ai released an abliterated GGUF of CyberStrike-OffSec-35B, a pentesting-tuned Qwen3.6 MoE model.
- The base CyberStrike-OffSec-35B is fine-tuned for structured tool calls in offensive-security agent harnesses.
- Base model benchmarks: SecEval 81.39%, MITRE ATT&CK 93.94%, CWE 93.05%, both #1 rankings.
- Abliteration removes the refusal direction from model weights without retraining or fine-tuning.
- Ships Q6_K (29.2 GB), Q8_0 (37.8 GB), and BF16 (71.1 GB) under Apache-2.0 license.
- Runs on llama.cpp, Ollama, LM Studio, and vLLM; supports 262K context window.
A new Hugging Face drop pushes local, agent-ready offensive security tooling one step further: Huihui-CyberStrike-OffSec-35B-abliterated-GGUF is a quantized, refusal-stripped variant of a pentesting-tuned Qwen model, packaged for llama.cpp, Ollama, LM Studio, and vLLM.
The upstream checkpoint is oyildirim/CyberStrike-OffSec-35B, described as an autonomous offensive-security agent fine-tuned on Qwen3.6-35B-A3B that emits structured tool calls with correct agent routing and clean termination, the behaviours a tool-calling harness actually needs. Huihui-ai then ran it through abliteration to strip refusals, and shipped GGUF quantizations that fit on a single workstation GPU.
Abliteration, briefly
Abliteration is a weight-editing trick, not a retrain. It removes the refusal direction from an open-weight LLM, producing a model that responds to prompts the original would refuse without further fine-tuning or system-prompt jailbreaks. The name splices ablation, the surgical removal of part of a system, with refusal.
The idea comes from Arditi et al.'s finding that refusal is mediated by a specific direction in the model's residual stream. The technique compares activations on harmless and harmful prompts to calculate that direction, then edits the weights so the model stops emitting refusals. Huihui's release uses the