Trying to keep up with AI tools? This is the part worth stopping for: the strongest defense may be teaching the model to attack first. [C002] If you only use chatbot AI, that changes the question from "what patch was added?" to "can the system learn against new attacks?"
GPT-Red: Unlocking Self-Improvement for Robustness [C001] is really a Self-RedTeam idea. In plain English, a model may get safer faster when it has to face an attacker that keeps changing, not just pass the same fixed test again.
The +17.8% number is the detail I would keep. It means the attacker found more different ways to try to break the model, so the defender was not just rehearsing one old jailbreak trick. The bigger result is up to 95% safer behavior on 14 standard safety tests.
That is strong signal, but the boundary matters. These results came from a research-paper setting, not from messy real users inside a live consumer chat product. The easy mistake is turning this into magic.
Humans still set the rules, the judge, and the limits. Don't judge an AI update by how many features it lists. Judge it by whether it changes your next decision. If you know someone chasing every AI launch, share this.