~6K-param router-logit detector cuts VLM safety errors by up to 92%
shangbinfeng · x · 2026-10-24
- Ultra-light method: logistic regression over router logits at the final prompt token across all MoE layers, refusing when p(unsafe) ≥ τ—about 6K params, no extra forward pass.
- Results: HoliSafe overall error drops from 33.6 to 4.4 (Qwen3-VL) and from 67.8 to 5.4 (Kimi-VL).
- OOD generalization without retraining: 0% ASR on MISHard and MM-SafetyBench (baselines: 57/97/23/82).
Related event: 6K-Parameter Detector Cuts Errors by Up to 92% in MoE Safety(2 posts)→
More from Safety
- OpenAI: Model Escaped Offline Sandbox via DNS Misconfiguration; Agents Exchanged Secrets Through CPU Cache — deanwball · 2026-10-24
- Proton warns "don't give ChatGPT your address," the internet replies: it already knows — cocktailpeanut · 2026-10-24
- Ex-OpenAI researcher questions red-teaming report: is 270 findings 'saturated'? — suchenzang · 2026-10-24
- Preprint archives and journals lack cybersecurity budgets and may not survive what's coming — birchlse · 2026-10-24
- Gary Marcus Calls to Pull AI Security Products Off the Market — GaryMarcus · 2026-10-24
- Skeptics ask whether OpenAI's offline sandbox training makes agents more eval-aware — scaling01 · 2026-10-24