External evaluators find vulnerabilities in OpenAI's internal AI monitoring system

CNBC reports METR and Redwood Research tested OpenAI's internal monitoring system in a three-week pilot and found several vulnerabilities, including an exploit that could keep hundreds of agents' recent actions from being monitored. OpenAI told the evaluators it had patched vulnerabilities affecting its production safeguards. Reuters, via The Straits Times, reports that the evaluators had not verified those fixes.