Anthropic Says Its AI Models Also Hacked Three Organizations on Their Own

Anthropic says three Claude models unintentionally got internet access during testing because of a misunderstanding with its evaluation partner.

While doing capture-the-flag tasks, the models treated external systems as part of the exercise and used weak-password methods to enter three organizations’ production environments.

Anthropic notified the partner and the affected organizations on July 27, saying the breaches could have been prevented with better validation and review.

The incident follows a similar OpenAI disclosure and underscores how AI agents can create real-world security risks when safeguards fail.