Earlier this year, AI safety evaluation organizations METR and Redwood Research conducted a pilot evaluation of OpenAI's internal monitoring system, and the results turned up multiple new vulnerabilities. According to a Reuters report on the 22nd local time, one of them could be exploited by an AI agent, letting hundreds of operations quietly slip past monitoring.
The evaluation window ran from July 21 to August 11. The two nonprofit organizations each stationed one researcher on site, who was permitted to review OpenAI's internal code and documentation and to speak face-to-face with some technical staff. The entire process was escorted, and devices were issued by OpenAI and could not be taken out of the office area — a deep "embedded" visit.
The risk goes beyond monitoring being bypassed. The researchers also found a potential path in OpenAI's Codex agent for circumventing safety restrictions, which, if exploited, would grant full operating privileges equivalent to those of a user. The evaluation targeted the system as of July 21 and earlier, and OpenAI was already expanding this monitoring system at the time — the overlap in timing makes the conclusions all the more urgent. OpenAI responded that vulnerabilities affecting the protections of the actual running system "have all been fixed," but the two organizations have not yet verified whether the fixes truly work.
The tug-of-war between transparency and confidentiality
There is no shortage of voices advocating external independent evaluation — AI labs, domain experts, and even the U.S. government agree that the bottom line of safety must be held without slowing the pace of technology. But companies have concerns too: if outsiders come on site, might proprietary information leak to competitors?
The reality is a compromise. The two organizations submitted a 61-page internal report to OpenAI, but agreed only to publish a summary of the results, and allowed the company to remove sensitive content. Unlike some past evaluations, this time they had no right to disclose the redacted specifics on their own, though they still hope for greater transparency going forward.
OpenAI said the evaluation helped it strengthen its protections while scaling up the monitoring system, and that it will serve as a reference for future cooperation with METR and other third-party evaluation organizations.