OpenAI releases GPT-Red to improve model robustness through automated red teaming
Original title:GPT-Red: Unlocking Self-Improvement for Robustness
AI summary
OpenAI trained an automated red-team model, GPT-Red, to find vulnerabilities before deployment and generate attacks during training to improve model robustness. GPT-Red can break almost all previous models, and its attacks were used in adversarial training for GPT-5.6 Sol, reducing the model's failure rate on direct prompt injection benchmarks to one-sixth that of the best production model four months earlier. GPT-Red was trained through self-play reinforcement learning, using an unprecedented scale of compute in OpenAI post-training.
Key Takeaways
- OpenAI trained an automated red-team model,GPT-Red,to find vulnerabilities before deployment and…
- GPT-Red can break almost all previous models,and its attacks were used in adversarial training for…
- 6 Sol,reducing the model's failure rate on direct prompt injection benchmarks to one-sixth that of…
- GPT-Red was trained through self-play reinforcement learning,using an unprecedented scale of…
Quick Facts
- Published
- Updated
- Source
- OpenAI:官网动态(RSS · 排除企业/客户案例)
- Category
- Model Releases
- AI score
- 67
OpenAI 用自博弈训练出的红队模型 GPT-Red,把直接提示注入攻击成功率压到了 0.05%,而且没有降低模型能力。做 AI 安全的人应该好好读一下他们怎么实现这个飞轮的。
Related entities
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: OpenAI releases GPT-Red to improve model robustness through automated red teaming. July 15, 2026. https://new.nat6.net/en/news/cmrmc527c0088bivcfrd3s2uz/