OpenAI just shipped a model that writes exploits and they called it alignment
GPT-6 Cyber can find zero-days and write working shell code. The alignment strategy is hoping it refuses when bad actors ask.
Read the note 1 min read
From the notebook
1 note on this topic.