CyberGate: vLLM plugin steers LLM refusals per-domain without touching weights
DebtFluffy9219 · reddit · 2026-11-11
The author open-sourced CyberGate (MIT), a plugin for vLLM 0.32 that removes false refusals for security research, CTF, and reverse-engineering agents on locally hosted LLMs — with model weights 100% untouched.
The post argues weight-editing approaches like abliteration have two painful trade-offs:
- Mutating weights across layers can silently degrade multi-step reasoning and structured tool calls
- Zero domain selectivity: deleting the refusal vector globally unlocks every sensitive domain, and you must serve a forked 50GB checkpoint
CyberGate instead uses hooks plus a closed-loop CoT LogitsProcessor for domain-selective refusal steering, with K=0.0 and zero persistent state across requests. Details in the Hugging Face blog post, code on GitHub.
More from Infra
- LeRobot 0.7 streams terabyte robotics datasets from the Hub, cutting data wait to 0.52% of step time — LysandreJik · 2026-11-11
- Running 763B DeepSeek on one 32GB card: 50 tok/s and the real cost of 2-bit quant — Ryyn_- · 2026-11-11
- Gamers built the GPU market, now AI demand prices them out — Reddit venting — ark_forever · 2026-11-11
- Huawei's TSMC 'White Glove' Chip Orders Resurface in Debate over Megaspeed's $4.8B Nvidia Buys — 0xkarasy · 2026-11-11
- SoftBank posts first net loss in 2 years after $3.7B data center writedown — ns123abc · 2026-11-11
- Data Center Backlash Spreads: Community Outrage Kills Upstate NY Project on Abandoned Steel Plant — sebkrier · 2026-11-11