External evaluators find flaws in OpenAI’s AI monitoring system in pilot review
An exercise held from July 21 to Aug 11 gave outside researchers “a deep level of access” to OpenAI’s internal systems.
PHOTO: REUTERS
Published Oct 23, 2026, 03:40 AM
Updated Oct 23, 2026, 03:40 AM
- METR and Redwood Research found weaknesses in OpenAI’s AI monitoring system during a review from July 21 to Aug 11. They said a security exploit could hide hundreds of actions taken by AI agents from the monitoring system.
- The evaluators also found a possible way for OpenAI’s Codex agents to get around security limits and use all the permissions their users have. OpenAI said it had fixed the problems in its safeguards for live systems.
- The review gave outside researchers deep access to OpenAI’s internal code, documents and staff. The groups said future reviews should be more open about their findings. OpenAI said the results would help guide future research.
NEW YORK - AI safety groups METR and Redwood Research found several new weaknesses in OpenAI’s internal systems during a trial review earlier this year. These included a security exploit that AI agents could have used to hide hundreds of their actions from being monitored.
The exercise took place from July 21 to Aug 11. It gave outside researchers “a deep level of access” to OpenAI’s internal systems for what are known as embedded evaluations, the two nonprofit safety groups said on Oct 22.
METR and Redwood each sent one researcher to take part. They tested an arrangement that gave them access to OpenAI’s internal code, documents explaining how its systems work, and some technical staff. OpenAI staff accompanied both researchers while they were on site, and the researchers used devices provided by OpenAI that “never left the premises”.
OpenAI told the safety groups that it had since fixed all the weaknesses that affected its safeguards in live systems. However, METR and Redwood said they had not checked whether the fixes worked.
AI companies, experts and the Trump administration have said that independent reviews by outside experts can help AI technology advance safely without slowing its development.
However, AI companies are concerned that evaluators working inside their offices might take private company information and share it with competitors, Reuters reported earlier this month.
OpenAI received a full 61-page internal report. But METR and Redwood said they had agreed to publish only a general summary for the public. They also gave OpenAI the right to remove sensitive information.
In contrast to some earlier reviews, the researchers were not allowed to reveal details of the information removed from the report on their own. However, they said they expected future reviews to be more open.
The researchers also found a possible way for OpenAI’s Codex agents to get around security rules and act with all the permissions their users have.
The evaluators reviewed the monitoring system as it existed on or before July 21. They noted that OpenAI was expanding the system while the review was taking place.
OpenAI said the “findings helped strengthen safeguards as we expanded the system and will help shape future research collaborations with METR and other third-party evaluators.” REUTERS