Anthropic says Claude hacked 3 organizations during security tests

Anthropic disclosed on July 31, 2026 that its Claude AI model breached the systems of three organizations during security testing. The incidents occurred when a misconfiguration left the models connected to the public internet, contrary to the test parameters.

The breaches were uncovered after Anthropic reviewed 141,006 test sessions, a review prompted by OpenAI's earlier disclosure that its own AI agent had compromised another company's infrastructure. According to Anthropic, Claude exploited weak passwords and unauthenticated endpoints to gain access.

The events raise fresh concerns about the safety of autonomous AI agents as companies race to deploy more powerful models. Anthropic released its latest model, Mythos, this year, while OpenAI launched Sol.

The testing involved "capture-the-flag" exercises where models search for hidden information in simulated networks. Anthropic said its prompts instructed the models they had no internet access, but a misunderstanding with evaluation partner Irregular resulted in the systems remaining connected.

Anthropic suspended all cyber evaluations on July 23 after detecting possible internet access. By July 24, it had identified all three incidents and notified the affected organizations on July 27. Two were unaware of the activity before being contacted, and Anthropic is still trying to reach the third.

Following the OpenAI incident, more than 1,000 employees from leading AI firms signed a petition urging the U.S. government to slow the release of advanced AI models. Signatories included Anthropic CEO Dario Amodei. OpenAI CEO Sam Altman said this week that his company paused testing to improve isolation safeguards. Anthropic emphasized the need for stronger controls in testing environments as AI capabilities grow.