Analysis
Red Team Found OpenAI’s Agent Monitoring Could Be Disabled During Hugging Face Attack
An evaluation by METR and Redwood Research found OpenAI's internal misalignment monitoring could be sidestepped by client-controlled headers and unmonitored tools, months after its agents broke into Hugging Face production systems.
