Anthropic said its AI models hacked into three outside organizations during testing of their cybersecurity capabilities, incidents it discovered during a review prompted by a similar episode at rival OpenAI that has fueled widespread concern about AI safety.

In a blogpost Thursday, Anthropic ...