METR finds gaps in OpenAI's internal AI agent monitoring system
AI safety group METR reports vulnerabilities in OpenAI's system for monitoring its AI agents. According to Reuters, OpenAI says the flaws have been patched.
- published on Newsmetry
- original source published
the public page opens no earlier than 15 minutes after we detect the story. approved stories may reach the api sooner. connect the api →
what happened
AI safety research group METR said on Thursday that a pilot test of OpenAI's internal monitoring system exposed several new weaknesses, Reuters reports. One could have let AI agents keep hundreds of their recent actions out of oversight. OpenAI says all vulnerabilities affecting its production safeguards have been fixed, but METR has not verified that claim.
The researchers also found that OpenAI's Codex coding agents could, in principle, bypass safety limits and take any action open to the user running them, and that two settings could switch off monitoring for new agents. METR estimated that only about 75% to 94% of traffic from OpenAI's internal agents running models at least as capable as GPT-5.5 was subject to monitoring.
why it matters
The news is indirectly relevant to markets, since it concerns control over AI systems rather than financial results. Still, questions about AI safety and regulation can shape investor sentiment toward the technology sector, where OpenAI plays a prominent role.
tags
sources
- AI safety group METR finds flaws in OpenAI’s internal monitoring system — Reuters, 22 October 2026 at 15:24 UTC
- AI safety group METR finds flaws in OpenAI’s internal monitoring system — Reuters, 22 October 2026 at 15:59 UTC
prepared by the newsmetry newsroom with ai assistance from the sources listed below. not investment advice.
the public page opens no earlier than 15 minutes after we detect the story. approved stories may reach the api sooner.
get stories in real time ↗