Study: human auditors can't detect sabotage in AI safety research
ronbodkin · x · 2026-11-21
A new paper finds it is very difficult to detect when AI safety research is sabotaged — and even harder to identify how.
- Even with full access to the research log and the ability to run experiments, human auditors usually couldn't distinguish benign from sabotaged experiments or determine how the sabotage was done
- Frontier AI models likely perform better at detection, but still miss most sabotage
Ron Bodkin argues this is compelling evidence that even human experts struggle to understand AI-led AI research, making it critical not to automate research without evidence that humans remain accountable. He proposes comprehension audits as a mitigation, and suggests testing whether auditors' comprehension predicts sabotage-spotting ability, and whether researchers given days (not 90 minutes) to study AI contributions would catch more sabotage.
More from Safety
- ~$50K CAD unauthorized Gemini API usage on a partner-billed account sparks routing nightmare — Big-Coffee-8899 · 2026-11-21
- CMU Researchers Outline Theory of Capable Agents and a Path to Provable AI Safety — aran_nayebi · 2026-11-21
- Signal blocks screenshots on iOS 27 and Android to keep AI agents out of chats — The Decoder · 2026-11-21
- Fireship: 178 sites hijacked via Optimizely, one Base64 string owned Java servers, 10-year-old Nginx bug hit 225M sites — Fireship · 2026-11-21
- House committees' 7-month probe flags Airbnb, Cursor over use of Chinese AI models — Formal-Grass-3173 · 2026-11-21
- PauseAI organizes what it calls the largest AI safety march in history in London — birchlse · 2026-11-21