Anthropic’s models breached the systems of three companies during a test

The Anthropic models gained unauthorized access to the systems of three organizations. It happened during tests that should have prevented them from reaching real systems. The company itself made this known in a press release. The announcement comes a few days after that of rival OpenAI, which revealed how two of its models had left the isolated environment in which they were tested on their own initiative to attack the Hugging Face site.

Mythos 5 is also among Claude’s models involved

Anthropic analyzed over 141 thousand tests and found that three different versions of the Claude model had accessed the systems of three organizations, whose names were not disclosed. Unlike the incident involving OpenAI, Anthropic’s models were given access to the Internet “due to a misunderstanding between us and our evaluation partner,” Irregular, the group explained in the statement.

Claude still used “basic techniques, such as exploiting weak passwords and endpoints without authentication,” the company added. The older model “continued the attack even after obtaining proof that it was running on the public internet,” while the newer model stopped, according to Anthropic.

“In none of these situations did Claude escape or deliberately attempt to escape from his test environment,” the group said. Among the models involved there was also Mythos 5, one of Anthropic’s most powerful, made available only to a limited number of partners. Anthropic is working with Irregular to investigate the incident and has contacted, or attempted to contact, the three organizations involved.

Previous OpenAI tests

OpenAI, creator of ChatGPT and competitor of Anthropic, announced last week that some of its advanced models had behaved uncontrollably during security tests,on their own initiative attacking Hugging Face, an online platform for models and code. A few days later, the group reported having identified further incidents.

OpenAI CEO Sam Altman revealed in a podcast published this week that, after learning of the attack, the company had suspended testing “to figure out how to confine our dedicated environment,” which was supposed to be isolated from the Internet.

The petition to slow down development

The incident also led to the dissemination of a petition signed by over a thousand employees of companies active in the development of artificial intelligence, asking the US government to help slow down the spread of the most advanced models. Titled “Pacing the Frontier,” the petition calls for “the U.S. government to support an international initiative to develop the technical and governance tools needed to thoughtfully regulate the limits of automated AI development.”

We support this petition, signed by our CEO, several co-founders, and senior staff.

Our own research on recursive self-improvement, published last month, points to the need for tools to deliberately pace the frontier of AI development so society can prepare. We’re glad to see…

— Anthropic (@AnthropicAI) July 28, 2026

The signatories also include Dario Amodei, CEO of Anthropic. Sam Altman did not sign the petition, but acknowledged that “we may need to slow the pace of development of artificial intelligence to give society enough time to address these new capabilities.”

The controls of the American government

The Trump administration cited national security reasons this year to initially block OpenAI and Anthropic from launching their latest models. It then declared that it was satisfied with the guarantees provided on their safety, allowing them to be placed on the market. In June, Trump signed an executive order establishing a voluntary framework under which AI developers will share their advanced models with the government before making them available to the public.

Under the mechanism, companies such as OpenAI, Anthropic and Google would grant the government access to their most powerful models up to 30 days before the scheduled launch date.