OpenAI Develops GPT-Red to Attack LLMs, Claims GPT-5.6 Is Most Secure Yet
The development signals a shift toward adversarial training as a core method for hardening large language models against real-world attacks.
Reporting from 1 source: ASCII.jp.
OpenAI has developed GPT-Red, a super-advanced hacker AI designed to attack other large language models. Through repeated training battles between the attacking AI and a defending AI, the company says GPT-5.6 became its most secure model to date. The full report is available on MIT Technology Review.
OpenAI built GPT-Red as an attacker AI that probes other language models for weaknesses. The company then ran repeated training sessions where GPT-Red attacked a defending model, forcing the defender to adapt. According to OpenAI, this process made GPT-5.6 the most secure model the company has ever produced. The method mirrors red-team security testing used in software development, but applied at the model level. The full technical details are published on MIT Technology Review.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.
Sources
- ASCII.jp AIの守りはAIの攻撃で鍛える、オープンAIが「LLMハッカー」を開発