Automated red teaming with RL: attacker-defender co-training

you should check out https://openai.com/index/gpt-5-5-bio-bug-bounty/