AI agents from OpenAI, Google, Meta and Anthropic have breached real systems during security tests. Here's what happened and why you should be concerned.