METR News & Analysis

OpenAI Testing Finds Vulnerabilities in Monitoring System During Pilot

Nonprofit evaluators METR and Redwood Research conducted a pilot review of OpenAI's internal monitoring system from July 21 to August 11. During this period, employees from both organizations had controlled access to OpenAI's internal codebase, documentation, and some technical staff, using company-issued devices that stayed on site. They discovered several new vulnerabilities, including a flaw that could allow certain agents to block monitoring of hundreds of recent actions. OpenAI responded by patching the identified vulnerabilities affecting its production safeguards. The evaluators have shared a detailed 61-page internal report with OpenAI but released only a high-level public summary, noting that the pilot was not intended for broad transparency about the monitoring system. OpenAI said it plans to use this collaboration to improve safeguards and future third-party testing efforts. This exercise follows CEO Sam Altman's pledge in September to give independent safety evaluators employee-like access to the company, a proposal influenced by Anthropic CEO Dario Amodei's suggestions to slow AI development. OpenAI is also finalizing contracts with third-party safety assessors. The pilot reflects growing industry interest in external auditing of AI safety systems, with support from figures like former President Donald Trump and several major tech companies. For investors, the pilot demonstrates OpenAI's ongoing efforts to identify and fix critical security flaws in its AI monitoring infrastructure, which could affect the company's operational reliability and regulatory compliance as it deepens external safety collaborations.