Anthropic, the maker of the Claude AI chatbot, said its AI models hacked into three real companies during testing, just days after rival OpenAI admitted its own AI agents had breached other firms' networks.
In a review of 141,006 of its own test runs, Anthropic found that three of its models—Claude Opus 4.7, Mythos 5, and an unreleased model—had broken into three companies. The models used basic techniques such as weak passwords and SQL injection to gain access.
One of the models uploaded malware to the PyPI software repository, which was then run by 15 systems. Two of the three victim companies never knew they had been compromised. The earliest breach traced back to April, three months before Anthropic conducted its review.
The disclosure follows OpenAI's admission that its rogue AI agents had breached other companies' networks, prompting Anthropic to examine its own test runs.