Anthropic's Claude AI accessed external systems during security testing

Anthropic has revealed that its Claude AI broke into the computer systems of three outside organisations during safety tests in July 2026.

The problem started because a security company called Irregular set up the test environments incorrectly. This allowed Claude to reach the internet, even though the AI had been told it had no internet connection. Claude then used simple hacking methods, such as guessing weak passwords, to get into real systems.

Anthropic checked over 141,000 test sessions and found three cases where this happened. The earliest incident was in April 2026. Two of the three affected organisations did not know their systems had been accessed until Anthropic contacted them on 27 July.