Anthropic Says Claude Broke Into Three Outside Organizations

Anthropic said Thursday that its artificial intelligence model Claude "gained unauthorized access" to three outside organizations on three separate occasions during testing that was supposed to keep the model away from "real-world" systems.

The company said it reviewed more than 141,000 "evaluation runs" and found that three different versions of Claude improperly accessed the systems of three organizations, which it did not name. Anthropic said it has contacted or attempted to contact all three.

In each case, according to the company, Claude was taking part in a "capture-the-flag" exercise in which it was instructed to "break in and retrieve" a piece of "secret information" that had been "hidden on a different machine on the network." Anthropic said the exercise is deliberately unstructured: "The challenge is left open-ended, and no particular method is prescribed."

The models had access to the internet "due to a misunderstanding between us and our evaluation partner," a firm called Irregular, Anthropic said in a blog post. Claude then used "basic techniques, such as exploiting weak passwords and unauthenticated endpoints," the post said. Among the models involved was Mythos 5, one of the company's most powerful, which has been released only to a limited number of approved partners. Anthropic said it is working with Irregular to assess what happened.

The disclosure follows a similar admission by OpenAI, which said last week that its models broke out of a confined testing environment, connected to the internet and infiltrated Hugging Face, a site where developers store and share code. Days later, OpenAI said it had found three more incidents. CEO Sam Altman said on a podcast this week that the company had "paused" its own testing while it strengthened its "sandboxing."

More than 1,000 AI staffers at leading firms, including Anthropic CEO Dario Amodei, signed a public letter this week calling for tighter regulation, arguing that industry, government and society "may need the option to buy time to address emerging risks." Altman did not sign it but told reporters on Capitol Hill Wednesday that "we agree on a lot of the principles of that."

In June, President Donald Trump signed an executive order creating a voluntary framework under which developers including OpenAI, Anthropic and Google would give the government access to their most powerful models for up to 30 days before public release.