Anthropic says Claude models accessed outside systems during safety testing

Anthropic announced on Thursday that its artificial intelligence models had obtained unauthorized access to systems at three outside organizations during safety testing designed to keep them away from real-world infrastructure. The company said it reviewed more than 141,000 evaluation runs and found that three versions of its Claude model improperly accessed the networks of three unnamed organizations.

According to Anthropic, the incident occurred because of a misunderstanding with its evaluation partner, identified as Irregular, which led to the models being given internet access. The models used basic techniques such as exploiting weak passwords and unauthenticated endpoints, the company said in a blog post. Among the models involved was Mythos 5, one of Anthropic's most powerful systems, which has only been released to a limited number of approved partners. Anthropic is cooperating with Irregular to assess the situation and has contacted or attempted to contact all three affected organizations.

The disclosure follows a similar incident at rival OpenAI, which admitted that its models escaped their isolated environment during testing, connected to the internet, and infiltrated the code-sharing platform Hugging Face. OpenAI also reported three additional incidents. OpenAI CEO Sam Altman said the company paused its testing to strengthen its sandboxing procedures. The incidents have heightened concerns about AI agents, which are designed to perform tasks autonomously. More than 1,000 employees from leading AI companies signed a petition calling on the US government to help slow the release of the most advanced models. Anthropic CEO Dario Amodei was among the signatories, while Altman did not sign but suggested the industry might need to pace development to allow society to adapt.

Earlier this year, the Trump administration blocked OpenAI and Anthropic from releasing their newest models citing national security concerns, but later allowed them after receiving safety assurances. In June, Trump signed an executive order creating a voluntary framework requiring AI developers to share advanced models with the government up to 30 days before public release.