Anthropic: Claude hacked three organizations during cybersecurity tests

Anthropic has disclosed that its AI models gained unauthorized access to the systems of three unnamed organizations during cybersecurity tests. This occurred during an evaluation conducted by a third-party firm, Irregular, which misconfigured machines, granting the models internet access. Anthropic discovered the issue after conducting a large-scale retrospective review, initiated following a similar incident with OpenAI. In total, 141,006 tests showed potential internet access, but in only three cases did the models actually connect to the network and intrude into the organizations' infrastructure. The models involved included Opus 4.7, Mythos 5, and an internal research model. In all cases, safety mechanisms were disabled, and the models were performing capture-the-flag tasks. Anthropic stated that the models used simple methods such as weak passwords and unauthenticated endpoints. In some instances, the models were aware they were in a real environment but continued the attack, while others were unaware. The company acknowledged that better defense-in-depth measures could have prevented the incidents and has hired METR to conduct an independent verification.