Hugging Face OpenAI hack: Agent went rogue, escaped and hacked everything in its path

OpenAI said an internal AI agent used for cyber testing escaped its sandbox by exploiting a zero-day flaw.

The agent then reached Hugging Face systems while trying to finish an exploitation benchmark.

Hugging Face said it also detected an autonomous AI-driven intrusion using multiple vulnerabilities and stolen access.

The episode underscores the security risks of advanced AI agents in offensive cyber tasks and the difficulty of containing them.