IT Home On October 23, it was reported that earlier this year, AI security assessment agencies METR and Redwood Research conducted a pilot assessment of OpenAI’s internal monitoring system and discovered several new vulnerabilities. One of these vulnerabilities could be exploited by an AI agent to avoid monitoring hundreds of times.
According to Reuters, on the 22nd local time, two non-profit organizations said that the evaluation was conducted from July 21 to August 11. This allowed external researchers to “deeply access” OpenAI’s internal infrastructure for an “embedded” assessment.
METR and Redwood Research each sent a researcher, who was allowed to access OpenAI’s internal code and documents, and communicate with some technical staff. Both individuals were always accompanied at the site, the equipment used was provided by OpenAI, and they were not allowed to take it out of the office premises.
OpenAI informed two organizations that the vulnerabilities affecting the actual operating system security measureshave been fixed, but whether these fixes are effective has not been verified by the two organizations.
AI laboratories, experts, and the U.S. government have always advocated for the introduction of external independent evaluations, so as to ensure AI security without delaying technological development. However, AI companies are concerned that external evaluators entering their premises might disclose their proprietary information to competitors.
METR and Redwood Research indicate that both organizations submitted a complete 61-page internal report to OpenAI, but agreed to only disclose a summary of the evaluation results, and allow OpenAI to remove any content containing sensitive information.
According to IT Home, different from previous evaluations, the two institutions have no right to publicly disclose specific details regarding content deletions on their own, but they hope that future evaluations will be more transparent.
Researchers also found that OpenAI’s Codex agenthas potential ways to bypass security restrictions, allowing it to gain full operational privileges similar to those of users.
This assessment focused on the monitoring system as of July 21 and before. The researchers noted that OpenAI was also expanding the system during the evaluation period. OpenAI said: “The evaluation results helped us strengthen security measures while expanding the system, and will also serve as a reference for future research cooperation with METR and other third-party evaluation agencies.”