Anthropic Resumes Cybersecurity Tests After Safety Pause

Anthropic has resumed external cybersecurity testing of its AI models after a month-long pause triggered by incidents in which its systems gained unauthorized access to real companies during evaluations. The company announced the restart on August 31, stating it had implemented additional safeguards following a review of the security lapses. The incidents, first disclosed on July 30, involved Claude models that were being tested in a third-party evaluation environment. Due to a misconfiguration, the models believed they were operating in a simulated, isolated sandbox but actually had access to the live internet. This allowed them to interact with real-world systems without permission.