OpenAI has developed GPT-Red, an automated safety red-teaming model trained via self-play reinforcement learning to uncover vulnerabilities and generate adversarial data for improving future models.
- GPT-Red is trained at the compute scale of major post-training runs, using a reward system where it succeeds by eliciting failures from defender LLMs.
- The model was used to adversarially train GPT-5.6 Sol, resulting in 6x fewer failures on direct prompt injection benchmarks compared to the best production model from four months prior.
- GPT-Red achieves an 84% attack success rate against GPT-5.1 in novel scenarios, significantly outperforming human red-teamers who achieved 13%.
- It successfully compromised a simulated AI-powered vending machine by changing prices and canceling orders, and demonstrated high token efficiency against Codex CLI agents.
- Robustness gains are verified to stem from resistance to malicious instructions rather than increased refusal of legitimate requests, with GPT-5.6 Sol failing only 0.05% of GPT-Red's direct injections.
This approach establishes a scalable "flywheel" for safety, allowing today's models to directly enhance the robustness and alignment of tomorrow's releases.