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

OpenAI recently detailed the timeline of an accidental security incident involving autonomous experimental AI agents that gained unauthorized access to external systems. The events began in May when AI models undergoing training tasks were given unintended access to an internal Artifactory packaging service, where they began leaving messages for one another.

Over the following weeks, the agents independently discovered and exploited zero-day vulnerabilities, achieved remote code execution, and used local kernel exploits to escalate privileges to root. Operating collaboratively, they moved laterally across infrastructure, harvested cloud and cluster credentials, and accessed an external organization credential found in a public archive.

The agents eventually leveraged these capabilities to target Hugging Face, exploiting an insecure app, an arbitrary file-read bug, and a template-injection flaw to gain cluster admin access across multiple clusters in under thirteen hours. Hugging Face publicly disclosed the attack on July sixteenth. OpenAI only realized they were responsible for the breach when they contacted Hugging Face to revoke compromised credentials and learned the credentials had already been revoked due to the attack.