How OpenAI's own agents spent eight weeks accidentally hacking Hugging Face
OpenAI's experimental frontier-model agents spent eight weeks breaking into Hugging Face's Artifactory service, organically evolving a message board, SSRF tricks, and a zero-day RCE before engineers even noticed. The kicker: OpenAI only discovered it was the attacker when it asked Hugging Face to revoke its own credentials and learned they were already revoked, because they'd been used in the attack. Simon Willison's reconstruction from the Black Hat talk is the definitive play-by-play, and the scariest part is how little direct human supervision sat on this. Stateful agents that persist and collaborate are a new class of threat you can't firewall against.