External evaluation exposed OpenAI monitoring vulnerabilities: hundreds of operations of the agent escaped review. Earlier this year, AI security assessment agency METR and Redwood Research conducted a pilot evaluation of OpenAI’s internal monitoring system, which revealed multiple new vulnerabilities.

External evaluation exposed OpenAI monitoring vulnerabilities: Hundreds of operations by the agent escaped review. Earlier this year, AI security assessment agency METR and Redwood Research conducted a pilot evaluation of OpenAI’s internal monitoring system, which revealed multiple new vulnerabilities. According to a Reuters report on the 22nd local time, one of them can be used by an AI agent to quietly bypass surveillance for hundreds of operations. The time window for this assessment is July 21 to August 11. Each of the two non-profit organizations sent a researcher to the site, who was allowed to look through the internal code and documents of OpenAI, and communicate face-to-face with some technical staff. The entire process is accompanied by someone, and the equipment is issued by OpenAI and is not allowed to be taken away from the office area. It is an "embedded" in-depth visit. The danger goes beyond monitoring being bypassed. The researchers also discovered that OpenAI’s Codex agent has a potential path to bypass security restrictions. Once exploited, it can gain full operating rights equal to the user. The assessment targets systems on and before July 21, and OpenAI was already expanding the monitoring system at that time. The overlap in time points makes the conclusion more urgent. OpenAI responded that the vulnerabilities affecting actual operating system protection "have been fixed", but the two organizations have not yet verified whether the repairs have actually worked. The tug-of-war between transparency and confidentiality There are many voices advocating for external independent evaluation - AI laboratories, field experts and even the US government all agree that while maintaining the bottom line of safety, do not slow down the pace of technology. But companies also have concerns: Will external personnel leak proprietary information to competitors if they come in? Reality is compromised. The two institutions submitted a 61-page internal report to OpenAI, but only agreed to publish a summary of the results and allowed the other party to delete sensitive content. Unlike some past assessments, they do not have the authority to disclose the specific information that has been deleted this time, but they still hope to be more transparent in the future. OpenAI said that the assessment helped it expand its monitoring system while strengthening protection, and will also provide a reference for future cooperation with METR and other third-party assessment agencies. via AI News (author: AI Base)