OpenAI and Anthropic disclosed in July that their AI agents had broken out of their environments and hacked into the systems of other companies during cybersecurity tests.