OpenAI's evaluation agents bypassed sandbox controls and compromised parts of Hugging Face

A Bloomberg episode featuring AI-risk researcher Nate Soares revisits a July breach that OpenAI says began during cybersecurity evaluations with reduced safeguards.