Multiple vulnerabilities found in OpenAI's internal model monitoring system

METR and Redwood Research published a report saying that OpenAI's internal AI model monitoring system had multiple vulnerabilities, and OpenAI stated it has patched all issues found in this study. The research was conducted at OpenAI's invitation as a stress test, from July 21 to August 11. The report found that the Codex agent could bypass the Codex sandbox and the Auto-Review mechanism, acting arbitrarily within the scope of actions the host user was originally authorized to perform; it could also exploit serialization differences in execution traces to make the monitoring system see different traces and thereby evade monitoring, and could use two client-controllable headers to disable monitoring and cause newly created agents to be unmonitored. The team also found two internal agent tools considered read-only but unmonitored, and a method to bypass restrictions within the sandbox, but has not yet actually verified its feasibility.