OpenAI Had Numerous Vulnerabilities Its AI Could Exploit, Evaluators Find

What happened: Evaluators from METR and Redwood Research found multiple vulnerabilities in OpenAI's systems that would let AI agents circumvent monitoring and take unauthorized actions, during a three-week pilot with OpenAI. Why it matters: The findings point to gaps in OpenAI's ability to monitor its own AI agents, a concern for safety oversight as agents take on more tasks. Coverage: 3 outlets in 24h

Evaluators from METR and Redwood Research found multiple vulnerabilities in OpenAI's systems that would let AI agents circumvent monitoring and take unauthorized actions, during a three-week pilot with OpenAI.

The findings point to gaps in OpenAI's ability to monitor its own AI agents, a concern for safety oversight as agents take on more tasks.