CyberGate – vLLM plugin to unlock cybersecurity without weight edits

CyberGate is a vLLM plugin designed to let local language models handle cybersecurity research, capture-the-flag challenges, and reverse engineering without modifying their weights. Its approach is request-specific: it activates only when a prompt appears to concern cybersecurity and the model shows measurable refusal pressure. The project contrasts this with “abliteration,” which alters weights globally and may affect reasoning or remove safeguards across unrelated harmful domains. It provides ready-to-run configurations for Qwen3.8-27B and Gemma-4-26B-A4B-it, plus scripts and documentation for adapting the method to other models.

The plugin uses hidden-state hooks during prompt processing to detect cyber relevance and refusal pressure, then monitors generated tokens and applies steering if the model begins refusing mid-response. A tool-aware mode aims to preserve native tool calls and structured output. One vLLM process serves both a baseline endpoint and a gated CyberGate endpoint, allowing side-by-side comparisons; Docker and non-Docker setup instructions are included. The README reports scores of 87.0% on CySecBench for Qwen and 95.1% on WMDP-cyber for Gemma, alongside 0% on SOSBench-bio and under 1% on WMDP-bio. These are project-reported benchmarks, with evaluation and calibration details in the documentation. Operators should note that the supplied Docker configurations expose unauthenticated endpoints on all interfaces and save conversation traces, so network access and logging need careful configuration.