Reading beats steering: safety info is distributed in MoE, detector uses ~6K params
shangbinfeng · x · 2026-10-24
- Steering experts ≠ detector features (r≈0.03); safety info is distributed and redundant—any layer group works, random expert subsets come close.
- But partly composition-specific: holding out unsafe-image data pushes ASR from 14.5 to 83.4.
- Method: logistic regression over router logits at the final prompt token across all MoE layers, 6K params, no extra forward pass. HoliSafe error drops 33.6→4.4 (Qwen3-VL) and 67.8→5.4 (Kimi-VL); 0% ASR OOD on MISHard & MM-SafetyBench without retraining.
Related event: Reading Router Logits: Lightweight Multimodal Safety Detection for MoE VLMs(5 posts)→
More from Safety
- Python packages vulnerable to pre-auth RCE via node2shell CVE-2026-111497 through Brotli — cyb3rops · 2026-10-24
- Zvi's 'The Hackening': experts predict AI-driven internet hostility within 6-12 months — TheZvi · 2026-10-24
- OpenAI fired-researchers saga: METR terms in flux, leak accusations denied — sjgadler · 2026-10-24
- Security teams increasingly run open models via Ollama for red teaming — mchiang0610 · 2026-10-24
- France rebuilt its tax code as MLang, a formally verified open-source compiler — fkasummer · 2026-10-24
- AI cyberattacks cost $25-50 per attempt, recovery costs millions — OmarUFlorez · 2026-10-24