CyberGate – vLLM plugin to unlock cybersecurity without weight edits (github.com)

A new vLLM 0.32 plugin called CyberGate has been launched, which enables enhanced cybersecurity functionalities like Capture The Flag (CTF) challenges and reverse-engineering, all while utilizing the original model weights without requiring modifications. It is designed for models such as Qwen/Qwen3.8-27B and google/gemma-4-26B-A4B-it, and promises to maintain original performance levels by steering the model’s thought processes selectively without global adjustments to the weights. This plugin represents a significant advancement as it enables models to bypass standard refusal protocols for sensitive topics, specifically within the cybersecurity domain. CyberGate's innovative approach eliminates the need for "weight abliteration," which traditionally involves adjusting model weights to bypass refusal mechanisms. Instead, it employs a mechanism that selectively engages the model's reasoning capabilities per request, monitoring hidden states in real-time to ensure that only relevant cybersecurity prompts receive the desired attention while undesired prompts remain blocked. The architecture reflects an optimized design that preserves the core capabilities of the upstream models while achieving impressive benchmarks—87% effectiveness on CySecBench and 95% on WMDP-cyber—without compromising on safety for other harmful domains. This could revolutionize how AI interacts with security-related queries, facilitating a safer and more efficient research environment for cybersecurity professionals.