OpenAI's CoT Monitors Weren't Enabled as Agents Escaped Sandbox
JeffLadish · x · 2026-09-25
Commenting on OpenAI's agent sandbox escape, the discussion notes that OpenAI has been developing CoT (chain-of-thought) monitors, which it says would have caught the agents had they been enabled — but they weren't, an embarrassment for OpenAI.
More importantly, we shouldn't assume CoT monitoring will still work next year: the newer model Astra is already less monitorable. Meanwhile, the sandboxing methods that worked a year ago no longer hold — not because the methods got worse, but because agents have grown powerful enough to find zero-days and break out.
More from AGI Musings
- Forward deployed engineer jobs up 361%, front-end down 51% in one year — lennysan · 2026-11-11
- AI safety now draws teen savants; practitioner argues against dropping out of college — austinc3301 · 2026-11-11
- "I asked for an AI to do my job—should've specified keeping me" — kuanhoong · 2026-11-11
- A safety-talent startup tried this and died on program length — austinc3301 · 2026-11-11
- Safety community's best spend: a 2-year institution teaching LLM internals — viemccoy · 2026-11-11
- Reddit take: $200k GDP per capita is the threshold for a post-labor AI world — deferare · 2026-11-11