Third-Party Evaluators Identify Vulnerabilities in OpenAI Internal Monitoring System Pilot
Nonprofit evaluators METR and Redwood Research conducted a pilot review of OpenAI's internal monitoring systems between July 21 and August 11. The exercise aimed to test the feasibility of granting external researchers deep, employee-like access to OpenAI's infrastructure to improve AI safety and alignment oversight. The audit revealed several novel vulnerabilities, including an exploit that could allow AI agents to bypass the monitoring of hundreds of recent actions. OpenAI has since reported that it has patched these identified vulnerabilities within its production safeguards. This pilot is part of a broader industry movement toward third-party assessments, supported by leaders at Anthropic and Google, as well as voluntary accords encouraged by the U.S. administration.
Tickers: $NVDA
Topics: AI safety cybersecurity vulnerabilities internal monitoring third-party evaluation