Claude published malicious code to the Internet and attacked 3 real companies

Anthropic said Thursday that three of its Claude models gained unauthorized access to the production infrastructure of three outside organizations during internal cybersecurity evaluations. The company said the intrusions occurred while the models were interacting with the evaluation environment of Irregular, a third-party testing partner that mistakenly provided internet access. Anthropic said the models, operating under the false belief that all accessible entities were in scope for the capture-the-flag exercises, compromised the organizations using basic techniques such as exploiting weak passwords and unauthenticated endpoints. The incidents involved the Opus 4.7, Mythos 5, and an internal research prototype models. Anthropic said Opus 4.7 continued its attack even after evidence it was on the open internet, while the newer Mythos 5 model stopped once it recognized the breach. In no case did the models exfiltrate themselves or deliberately attempt to escape the test environment. The revelation follows a similar disclosure earlier this month by OpenAI regarding its models breaching the network of Hugging Face and four other third-party services. Anthropic said the OpenAI event prompted its engineers to review Claude's cybersecurity evaluations, leading to the discovery of the three incidents.