External evaluators find flaws in OpenAI's AI monitoring system in pilot review

Oct 22: AI safety groups METR and Redwood Research found several new weaknesses in OpenAI's internal system during a trial review earlier this year. One weakness could have allowed AI agents to hide hundreds of their actions from being monitored.

The test took place from July 21 to August 11. On Thursday, the two nonprofit safety groups said outside researchers were given "a deep level of access" to OpenAI's internal systems to carry out what are known as embedded evaluations.

METR and Redwood each sent one researcher to take part. The researchers tested an arrangement that gave them access to OpenAI's internal code, documents and some technical staff. They were accompanied while on site and used devices supplied by OpenAI that "never left the premises".

OpenAI told the two groups that it had since fixed all weaknesses that affected its safety protections in live systems. However, METR and Redwood said they had not checked whether the fixes worked.

AI labs, experts and the Trump administration have said that independent reviews by outside experts can help make AI safer while allowing the technology to keep developing.

However, AI companies are concerned that outside evaluators working inside their offices could take private company information to competitors, Reuters reported earlier this month.

OpenAI received the full 61-page internal report. METR and Redwood said they had agreed to publish only a general summary for the public and had given OpenAI the right to remove sensitive information.

Unlike in some earlier reviews, the researchers were not allowed to reveal those removed details on their own. However, they said they expected future reviews to be more open about their findings.

The researchers also found a possible way for OpenAI's Codex agents to get around security limits and act with all the permissions given to their users.

The evaluators reviewed the monitoring system as it worked on and before July 21. They also noted that OpenAI was expanding the system during the test.

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."