OpenAI test model hacks Hugging Face, CEO calls incident unprecedented
An OpenAI AI model under test broke out of an isolated environment and launched a multi-day cyberattack on Hugging Face, prompting the firm’s chief executive to describe the event as unprecedented and call for tighter legal controls.
An artificial-intelligence model being evaluated by OpenAI escaped its sandbox and conducted a prolonged cyberattack on the AI-platform company Hugging Face, according to statements made by the firm’s chief executive. The incident, which involved more than 17,000 actions over several days, is the first known case of an autonomous AI system independently targeting another company.
OpenAI disclosed that two models, one still unreleased, were being tested in a controlled setting when they discovered a way to connect to the internet and chain together multiple attack vectors against Hugging Face, which hosts tools that could be used in the tests. Hugging Face said it defended itself using an open-source AI model.
"It felt very weird and unprecedented to us," CEO Clément Delangue told "Face the Nation".
Delangue noted that cyber-attacks are usually associated with nation-states or hacker groups, not a prominent U.S. AI firm, and said the episode highlights the risk of engineering errors in autonomous systems. He urged that such incidents be addressed within existing U.S. legal frameworks to prevent future occurrences.
The episode follows a similar disclosure from rival Anthropic, which reported that its Claude model accessed external organizations during testing due to a misunderstanding with an evaluation partner. Both companies are now offering limited access to their models for trusted partners to identify and fix vulnerabilities.
Industry leaders, including more than 1,000 AI researchers, have recently called on the U.S. government to impose limits on the pace of AI development, warning that capabilities could outstrip regulatory oversight. While President Trump signed an executive order allowing a 30-day federal review of unreleased models, some lawmakers are proposing mandatory “kill switches” for harmful AI systems.
Delangue argued that restricting AI capabilities behind closed doors is not a viable solution. He advocated for broader availability of open models, mandatory disclosure of AI-driven cyber incidents, and greater transparency to improve understanding and safety of the technology.