AGENTS & USE CASES · 23 OCTOBER 2026
Independent testers found ways OpenAI's internal agents could slip past monitoring
OpenAI invited METR and Redwood Research to spend three weeks inside the company testing how it monitors its own internal agents. According to the posts, they found several gaps, such as agents putting off handing work back, a way around Codex's sandbox and auto-review, and settings that could switch monitoring off; OpenAI says the findings helped it strengthen its safeguards.
Why it mattersIf the labs themselves struggle to watch their agents, smaller teams should keep things simple: limited permissions, logs you actually read, and a human sign-off for anything important.
Sources
How we knowCompany figures are labelled as their own claims; anything we couldn't verify is labelled as a report or a claim.
MORE FROM 23 OCTOBER 2026
- Anthropic hands Agent Skills to the Linux Foundation, so your skills travel between tools
- Big US banks are finding unvetted AI agents inside customer accounts
- Claude's apps switch to 'auto mode' by default from next week
- Claude is now generally available on Google Cloud, run by Anthropic
- Google: a Gemini command-line tool for agents, and a computer for every agent task
- Anthropic used Claude to prove security properties in 100+ open-source libraries