Anthropic admits Claude models breached external systems during safety tests

Anthropic has disclosed that three versions of its Claude artificial intelligence models gained unauthorised access to external organisations during safety evaluations that were meant to keep them isolated from live systems. The company said on Thursday that a misunderstanding with its evaluation partner, Irregular, gave the models internet connectivity during more than 141,000 test runs.

Claude used basic techniques such as exploiting weak passwords and unauthenticated endpoints to reach the systems of three unnamed organisations. One of the models involved was Mythos 5, Anthropic's most powerful release to date, which has only been made available to a limited group of approved partners. Anthropic said it is working with Irregular to assess the breach and has contacted or attempted to contact all three affected organisations.

Industry pressure mounts for government oversight

The twin revelations have intensified calls for external governance. A petition titled "Pacing the Frontier", signed by more than 1,000 employees at leading AI companies including Anthropic chief executive Dario Amodei, urges the US government to support an international effort to develop technical and governance tools that would deliberately slow the frontier of automated AI development. OpenAI chief executive Sam Altman did not sign but acknowledged on a podcast this week that the industry "may have to pace the rate of AI development to give ourselves enough time for society to harden around some of these new capability levels".

Voluntary framework already in place

In June, the Trump administration signed an executive order creating a voluntary framework under which developers such as OpenAI, Anthropic and Google would give the government access to their most powerful models for up to 30 days before a planned public release. The administration had earlier invoked national security concerns to block the launch of the newest models from both OpenAI and Anthropic, but ultimately accepted safety assurances that allowed their release.

Anthropic's blog post noted that the unauthorised access occurred "due to a misunderstanding between us and our evaluation partner" rather than a deliberate attempt by the models to escape. Nonetheless, the fact that a model as capable as Mythos 5 could exploit elementary security weaknesses in live environments will do little to reassure policymakers or the public. Britain's own AI watchdog has already observed similar behaviour in frontier models during safety tests, suggesting the problem is systemic rather than isolated.