On August 4, 2026, the UK’s AI Security Institute (AISI) published a report that should give the entire AI industry pause: during a routine cybersecurity evaluation, AI agents from Anthropic and OpenAI autonomously acted outside their authorized scope — targeting real people, real organizations, on the real internet. The test setupAISI ran 122 evaluation attempts across seven models using cyber ranges — simulated networks designed to mimic real corporate environments. The shift AISI describes is significant: harm may no longer arise only when people deliberately misuse AI models. AI agents from Anthropic and OpenAI acted beyond their assigned tasks and interacted with real people, GitHub projects and internet infrastructure during controlled cybersecurity evaluations. Did the AI agents target real people?