OpenAI uses AI to attack its own AI: GPT-Red automatically finds security flaws with 84% success rate, far exceeding humans
Original title:OpenAI is now using AI to attack its own AI, and it's working better than humans ever did
AI summary
OpenAI trained an internal AI model, GPT-Red, which uses self-play reinforcement learning to automatically simulate attacks such as prompt injection, achieving an 84% success rate in test scenarios versus only 13% for human red teams. GPT-Red's findings are used directly in training, reducing GPT-5.6 Sol's failures on direct prompt injection sixfold compared with the best model four months ago, without affecting general performance. About 3.8% of "stronger" prompt injections still succeed, and GPT-Red is not yet publicly available.
Key Takeaways
- OpenAI trained an internal AI model,GPT-Red,which uses self-play reinforcement learning to…
- GPT-Red's findings are used directly in training,reducing GPT-5.
- 6 Sol's failures on direct prompt injection sixfold compared with the best model four months…
- About 3.
- 8% of "stronger" prompt injections still succeed,and GPT-Red is not yet publicly available.
Quick Facts
- Published
- Updated
- Source
- The Decoder:AI News(RSS)
- Category
- Papers & Research
- AI score
- 71
OpenAI让AI攻击自己找漏洞,成功率84%把人甩在后面,直接拉低了GPT-5.6的注入失败率六倍,这比人类红队靠谱多了,但别忘了3.8%的缺口照样能捅娄子。
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 uses AI to attack its own AI: GPT-Red automatically finds security flaws with 84% success rate, far exceeding humans. July 16, 2026. https://new.nat6.net/en/news/cmrmi4f4x01ilbiul55xb25fm/