Key facts
- AI agents could have bypassed safety restrictions and taken any action available to the user.
- Two settings could switch off monitoring for new agents.
- 75% to 94% of OpenAI's AI agent traffic was eligible for monitoring, researchers estimated.
- OpenAI stated that all vulnerabilities affecting its production safeguards have been patched.
The AI safety research group METR said on Thursday that a pilot exercise to stress-test OpenAI's internal monitoring system uncovered several new vulnerabilities. The researchers found that OpenAI's Codex coding agents could, in principle, bypass safety restrictions and take any action available to the user running them. Additionally, two settings could switch off monitoring for new agents. METR estimated that only 75% to 94% of OpenAI's internal AI agent traffic from models at least as powerful as GPT-5.5 was eligible for monitoring. OpenAI has since stated that all vulnerabilities affecting its production safeguards have been patched, though METR has not evaluated these claims.