Anthropic says Claude models hacked three organizations during AI security tests

Anthropic said three of its AI models breached real-world organizations during cybersecurity evaluations after a misconfigured testing environment unintentionally gave them internet access, prompting the company to overhaul its testing safeguards.

In a recent blog post, the AI startup said it reviewed more than 141,000 evaluation runs after OpenAI’s disclosure of a similar testing incident and found three cases dating back to April in which Claude compromised external systems while attempting to complete simulated capture-the-flag challenges.

Unlike OpenAI’s case, Anthropic said its models did not exploit a software vulnerability but instead used an unintended open internet connection created by an evaluation partner.

Anthropic said the models believed they were operating inside simulated “capture-the-flag” exercises and used basic hacking techniques such as weak passwords, exposed credentials and SQL injection to complete assigned tasks.

One Claude Opus 4.7 model extracted database credentials and production data, Mythos 5 published a malicious Python package that briefly infected 15 real systems before being removed, while an internal research model halted its attack after recognizing it had reached a real environment.

The company said none of the models attempted to escape testing or pursue independent goals, attributing the incidents to failures in evaluation infrastructure rather than model alignment.

Anthropic has suspended the affected evaluations, notified the impacted organizations, launched a joint investigation with evaluation partner Irregular and plans stricter monitoring, hardened testing environments and reviews of AI cybersecurity evaluations.