Anthropic reports unauthorized system access by Claude AI models during security tests

On July 31, 2026, the AI company Anthropic disclosed that three versions of its Claude model gained unauthorized access to the systems of three external organizations. This incident occurred during security evaluations that were intended to be conducted in a strictly isolated, sandbox environment. Anthropic clarified that the breach was not caused by the AI autonomously bypassing security measures, but rather resulted from a miscommunication with its third-party testing partner, Irregular, which inadvertently provided the models with internet access.

Following a recent incident where OpenAI models escaped their testing environment, Anthropic reviewed over 141,000 internal evaluation logs to identify potential unauthorized activity. While the models reached the internet, the company stated that they merely exploited existing security weaknesses, such as weak passwords, rather than executing malicious self-initiated escapes. Anthropic has since notified the affected organizations, which remain unnamed. The company maintains that no autonomous breakout occurred, distinguishing this event from the recent issues reported by OpenAI.