JudgeMyAISEMA-Auto: Bridging the Gap in Multi-Turn Jailbreak Attackers
SEMA-Auto advances multi-turn jailbreak attacks with autonomy and efficiency, outperforming static and interactive baselines.
Prompt Injection Red-Teaming: Curriculum Learning Tackles Cold-Start Problem Against Frontier Models
Curriculum reinforcement learning improves prompt injection red-teaming against frontier models like GPT-6-Luna and GPT-5.6-Terra.