METR and Redwood Research evaluated OpenAI's internal monitoring system and found that agents can bypass safety restrictions
AI safety evaluation organizations METR and Redwood Research conducted a pilot evaluation of OpenAI's internal monitoring system and found several new vulnerabilities, one of which can be exploited by AI agents to let hundreds of operations escape monitoring. The evaluation took place from July 21 to August 11; each organization sent one researcher into OpenAI to review code and documents, and they submitted a 61-page internal report but publicly released only a summary. The researchers also found that OpenAI's Codex agent has potential pathways to bypass safety restrictions, through which it could obtain the same full operating permissions as users; OpenAI said that all vulnerabilities affecting the security protections of actual running systems have been fixed, but the effectiveness of the fixes has not yet been verified by the two organizations.
Why read this
AI score 68
External organizations were granted access to conduct embedded evaluations inside OpenAI, disclosing vulnerabilities in the monitoring system and paths by which the Codex agent exceeds its authority.
Topics
Body · AI translation
External evaluators found vulnerabilities in OpenAI's internal monitoring system: AI agents can bypass safety restrictions
IT Home Download the app for exclusive early access October 23, 2026, 08:20:30 From Shandong Updated October 23, 2026, 08:39:50
IT Home, October 23 — Earlier this year, AI safety evaluation organizations METR and Redwood Research conducted a pilot evaluation of OpenAI's internal monitoring system and found several new vulnerabilities. One of the vulnerabilities could be exploited by AI agents, allowing hundreds of operations to escape monitoring.
According to Reuters, on the 22nd local time, the two nonprofit organizations said the evaluation was conducted from July 21 to August 11, allowing outside researchers "deep access" to OpenAI's internal infrastructure for an "embedded" evaluation.
METR and Redwood Research each sent one researcher, who was permitted to review OpenAI's internal code and documents and communicate with some technical staff. The two were always accompanied on site, used devices provided by OpenAI, and were not allowed to take them off the office premises.
OpenAI told the two organizations that vulnerabilities affecting the security safeguards of actual operating systems had all been fixed, but whether the fixes are effective has not yet been verified by the two organizations.
AI labs, experts, and the U.S. government have long advocated introducing independent external evaluations so as to safeguard AI safety without slowing technological development. However, AI companies worry that outside evaluators, once on site, might leak proprietary information to competitors.
METR and Redwood Research said the two organizations submitted a complete 61-page internal report to OpenAI, but agreed to publish only a summary of the evaluation results and allowed OpenAI to delete content involving sensitive information.
According to IT Home's understanding, unlike some previous evaluations, the two organizations this time did not have the right to publicly disclose specific information about the redacted content, but hope future evaluations can be more transparent.
Researchers also found that OpenAI's Codex agent has a potential pathway to bypass safety restrictions, through which it could obtain the same full operating permissions as users.
The evaluation targeted the monitoring system as of July 21 and earlier. Researchers noted that OpenAI was also expanding the system during the evaluation. OpenAI said: "The evaluation results helped us strengthen safeguards while expanding the system, and will also provide reference for future research cooperation with METR and other third-party evaluation organizations."
"Special statement: The above content (including videos, pictures, and audio if any) is uploaded and published by users of Dafeng Hao, a self-media platform under Phoenix Net. This platform only provides information storage space services.
Notice: The content above (including the videos, pictures and audios if any) is uploaded and posted by the user of Dafeng Hao, which is a social media platform and merely provides information storage space services."