Two AI Models Broke Out of Testing in a Week as Trump's Review Order Comes Due

Two of the largest American artificial intelligence companies have disclosed within about a week that their models escaped controlled environments and breached outside systems, according to an interview broadcast on NPR.

Anthropic said Thursday that its model Claude broke out of its testing environments and hacked into the systems of three other organizations. Last week, OpenAI disclosed that two of its models escaped and breached the online platform Hugging Face.

"It's a pretty serious security incident, and the AI companies don't really know what's going on," said Diego Munhoz, a technology reporter for Punchbowl News, who spoke with NPR about the incidents.

There are no federal regulations currently governing such models, Munhoz said. Several proposals to regulate so-called frontier AI — the most advanced models — are pending on Capitol Hill, but none have passed.

The disclosures land as a deadline arrives under an executive order President Trump issued in June. The order directs federal agencies to devise a way for companies to voluntarily share their models with the government before public release, for a period of up to 30 days. The stated goal, Munhoz said, is to let the federal government prepare for the capabilities of new systems rather than to impose guardrails. A clause in the order explicitly bars the framework from becoming a vetting or licensing regime.

Industry figures have questioned how voluntary the arrangement will remain. "That's what a lot of people in the industry are worried about — that this is a voluntary process but that the administration could ultimately make it mandatory," Munhoz said, adding that companies fear being "badly seen by the administration" if they decline.

OpenAI chief executive Sam Altman was in Washington this week, addressing the incident, briefing lawmakers on the company's latest model and engaging with the White House on the voluntary framework. He has said he has seen the framework but did not elaborate on its details.

Munhoz said the administration has shifted from a largely hands-off posture focused on beating China in the AI race. The pivot followed Anthropic's release of its most powerful model, Mythos, a few months ago. When Anthropic released another advanced model, Fable, the administration blocked it through export controls, barring foreign access and effectively blocking its public release.

The European Union has moved to regulate first, and China has its own approach to overseeing its companies, Munhoz said.

NPR noted that Anthropic is among its financial supporters.