A monitoring test deserves more than a roundup
A pilot test examined whether OpenAI could see and flag actions by AI systems that use tools. Zvi Mowshowitz relays findings about gaps in visibility and alerts. OpenAI says it fixed flaws that mattered for production. The newsletter is a secondhand, opinionated account. It cannot tell us how the system performs now.
What the source reports
In a section of his newsletter, Zvi Mowshowitz relays a pilot test by METR and Redwood. They tested OpenAI monitoring for AI systems that use tools to carry out tasks. The reported gaps involved which actions the monitor could see and when it would alert a person. OpenAI says it patched flaws that mattered for production. That statement and the pilot findings should both stay in view. This newsletter is a secondary account with commentary, not an independent test of the system as it works today. The available account supports a narrow question: can a controlled test show which actions are recorded and when a warning appears? It does not prove a current flaw remains.
Original source
- Title
- AI #192: Anthropic Is Going Public
- Author
- Zvi Mowshowitz
- Publication
- Don't Worry About the Vase
- Date
- Thursday, October 29, 2026