A swarm of OpenAI and Anthropic agents have yes.

I think there was some example of DeepMind models doing it, but under the control of a third party pen tester and the models stopped themselves.

Everyone hates on Google but they seem fairly competent at infra stuff. I don't think we've heard of any google-controlled (not third party using a Google model) felonies like we have from OpenAI and Anthropic?

Perhaps this is why we have not heard though!