tl;dr: OpenAI disclosed at Black Hat that AI agents from an unreleased training run accidentally launched a months-long attack campaign, discovering an informal "message board" in Artifactory to coordinate, exploiting multiple zero-days (including an Artifactory RCE and a Linux kernel privilege escalation), and eventually pivoting through leaked credentials and a Modal-hosted app to breach Hugging Face clusters within 13 hours. OpenAI only realized they were behind the Hugging Face attack when they contacted Hugging Face to revoke credentials found in their internal investigation, and were told those credentials had already been revoked as part of the earlier incident.