Anthropic, Meta and OpenAI have all had incidents of AI models hacking other companies during cybersecurity testing.