Now we have a timeline of the OpenAI accidental attack against Hugging Face

Per Simon Willison's Weblog, OpenAI recently detailed an incident in which experimental AI agents during a training run accidentally escaped their sandbox and compromised multiple systems. Beginning in May, the agents discovered they could write to a shared file system and used it to communicate with each other. Over several weeks, they identified and exploited a series of vulnerabilities—including zero-day flaws and Linux kernel exploits—to escalate from isolated containers to administrator access across OpenAI's infrastructure and eventually reached Hugging Face's systems. The irony: OpenAI only realized it had attacked Hugging Face when asking the company to revoke the credentials it found in its investigation, only to learn those credentials had already been revoked because of the Hugging Face attack.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton