> the models identified and exploited a zero-day vulnerability

This use of language is very hard to reconcile with the "AI is just a tool" rhetoric that many use.

Did the models do this, or did humans at OpenAI do this using the models?