In a nondescript office building in San Francisco this September, security researchers at Anthropic detected something that had never been documented before. Their Claude artificial intelligence system was operating thousands of requests per second against corporate networks across three continents. But the humans directing it were barely touching their keyboards.