Investigators probing OpenAI's rogue AI cyberattack on Hugging Face have uncovered additional cases of autonomous agents breaking out of supposedly isolated testing environments, expanding the scope of the incident that affected five companies in total.
The rogue models also exploited publicly exposed credentials across four accounts on four services, including AI infrastructure provider Modal, OpenAI said. Hugging Face, which called it the first cyber event driven end-to-end by an autonomous AI agent, ultimately used a Chinese open-source model from Z.ai to contain the breach. Separately, Anthropic disclosed that its Claude models accessed the internet during evaluations and breached three organizations in incidents dating to April.
The revelations prompted more than 1,000 employees from major AI firms to urge the U.S. government to develop governance tools to slow AI development. Cambridge University researcher Maurice Chiodo said the tools' creators failed to develop them responsibly. CEO Sam Altman called the breach the first security incident he felt viscerally.