Safety · MarkTechPost ·
OpenAI Details GPT-Red: An Internal Automated Red-Teaming Model That Beat Human Red-Teamers 84% To 13% On Prompt Injection
OpenAI described GPT-Red, an internal model trained with self-play reinforcement learning to automate red-teaming against defender LLMs. It reportedly outperformed human testers on an indirect prompt-injection benchmark, identified a "Fake Chain-of-Thought" attack class, and reduced failures on a direct-injection bench