GPT-Red: Automated Red Teaming via Self-Play at Scale
GPT-Red is presented as an automated attacker used to adversarially train GPT-5.6 against prompt injections.
The paper says GPT-Red is trained through self-play against a changing set of defender agents in realistic red-teaming environments. The authors describe the run as the largest documented LLM safety training run, at a scale comparable to major RL post-training. They report that it broke models through GPT-5.5, found more successful attacks than human red-teamers, and generalized beyond its training setup. Source: HF Daily Papers' note.
The paper says GPT-Red is trained through self-play against a changing set of defender agents in realistic red-teaming environments. The authors describe the run as the largest documented LLM safety training run, at a scale comparable to major RL post-training. They report that it broke models through GPT-5.5, found more successful attacks than human red-teamers, and generalized beyond its training setup. Source: HF Daily Papers' note.
score 7