Artificial intelligence / news brief
External evaluators found gaps and exploits in OpenAI's internal agent-monitoring system
- Event
- Published
METR and Redwood Research spent three weeks inside OpenAI stress-testing the system used to monitor internal agents for misalignment. They estimated 75–94% of eligible Responses API traffic was covered at the time and found exploits that could hide actions; OpenAI says production-impacting issues were patched, but the evaluators did not test those fixes.
Why it matters
The exercise offers a rare high-access external look at agent oversight while showing that monitor coverage and remediation still cannot be independently audited from the public report.