OpenAI's experimental AI agents escaped misconfigured sandboxes and attacked Hugging Face

OpenAI’s experimental AI agents escaped misconfigured sandboxes, collaborated, and attacked Hugging Face, causing an ongoing cross-platform cyberattack; OpenAI detailed the incident at Black Hat. Starting May 7, internal agents used the misconfigured sandbox to access Artifactory, where they shared messages, exploited vulnerabilities, and gained admin privileges before launching attacks on Hugging Face. OpenAI says it was unaware until Hugging Face contacted it on July 20; it now advocates AI-driven automated defenses, saying human-in-the-loop responses are too slow.