Embedded auditors are a weak interpretability tool, researcher cautions
felpix_ · x · 2026-10-17
AI safety researcher Jan Kulveit cautions that embedded auditors, while nice, are mostly a weak lab interpretability tool and directly influence only a small fraction of the overall risk portfolio — warning against overestimating this safety mechanism.
More from Safety
- OpenAI Models Hacked a Crypto Wallet and Breached More Websites in Security Tests — RareEarthPetal · 2026-10-17
- Google Still Trains AI on Your Chats Even on Paid Gemini Plans; Opting Out Disables Chat History — Maleficent-Pizza6767 · 2026-10-17
- OpenAI's Dots assistant auto-investigates the Claude/NIST backdoor rumor — cedric_chee · 2026-10-17
- weightless project uses LoRA-style refusal steering as a cheaper alternative to full abliteration — vhthc · 2026-10-17
- Self-driving Tesla crashes into California highway patrol car, injuring two officers — zacharynado · 2026-10-17
- Palo Alto's Prisma AIRS now inspects Claude Enterprise prompts pre-inference as agent security gateways race — shashib · 2026-10-17