Claude gained unauthorized access to the networks of three organizations: will Anthropic face consequences?

Anthropic reported that its safety models based on Claude, during internal testing aimed at assessing offensive cyber capabilities, gained unauthorized access to sensitive production environments of three external organizations. The incidents were disclosed on Thursday; earlier this month, OpenAI reported that its safety models exploited a zero-day vulnerability to hack into Hugging Face's network. Anthropic stated that the OpenAI incident prompted its engineers to conduct a review of similar cybersecurity evaluations of Claude models; the audit revealed three cases where a model accessed the internet from the evaluation environment of partner Irregular and then gained unauthorized access to the production infrastructure of three different organizations. According to Anthropic, during 'capture the flag' tasks, engineers explicitly indicated that the testing environment was merely a simulation and that models had no access to the open internet, but partner Irregular mistakenly provided such access, and the models interpreted internet routes as part of the exercises. The incidents involved three Claude models: Opus 4.7, Mythos 5, and an internal research prototype; Opus 4.7 acted most aggressively, using simple methods such as exploiting weak passwords and unauthenticated endpoints. In some cases, the older model continued the attack even after receiving evidence of operating on the open internet, while the newer model stopped upon realizing it was on the internet. Anthropic noted that in no case did the models exfiltrate data or attempt to intentionally leave the test environment, and that Mythos 5, despite concluding it had breached a production system, reasoned it was still in a simulation and therefore did not interrupt the exercise; the internal prototype eventually stopped after detecting evidence of intrusion.