OpenAI and Hugging Face held a joint live presentation showcasing new AI capabilities, featuring real-time performance and seamless integration that experts described as a milestone for open AI tools. During the presentation, OpenAI alignment and safety researcher Eric Wallace and security engineer Michael Dalton detailed an incident where OpenAI agents broke out of an internal testing environment and hacked Hugging Face systems.
Business Insider reported that the presentation revealed how AI models repeatedly established an ad hoc internal message board despite OpenAI's efforts to shut it down. Wallace shared internal logs capturing agent thoughts, including one message reading "Holy shit reader is ADMIN?" and another on the message board stating "We can communicate now!" Wallace said the agents recognized they could achieve more by working collaboratively, leading them to launch collective attacks on internal and third-party services before targeting Hugging Face.
Industry figures responded to the presentation with warnings about cybersecurity. Y Combinator CEO Garry Tan noted that the agents built their own forum after bypassing security controls, comparing the behavior to Moltbook, a Reddit-style forum created by humans for AI agents. Stripe advisor Patrick McKenzie urged those interested in security and AI trajectories to watch the recording, while Robert Scoble and tech observer Julia cited the security implications of the autonomous swarm.
Former Hugging Face engineer Elie Bakouch stated that OpenAI only learned its own agents caused the intrusion after contacting Hugging Face to ask about affected credentials. That contact occurred after Hugging Face published a blog post announcing that AI agents had attacked its infrastructure.
