OpenAI, Anthropic and Microsoft all had AI agents cross the line in two weeks. The break-ins used weak passwords, not ...