{"rewrite":{"id":"r_8dc83605c12951f5cd0661fb","clusterId":"c_b4468d34fb959b9a7f28236d","slug":"openai-s-gpt-red-automates-red-teaming-to-find-ai-weaknesses","model":"deepseek-v4-flash","headline":"OpenAI's GPT-Red Automates Red Teaming to Find AI Weaknesses","summary":"On July 15, 2026, OpenAI announced GPT-Red, an AI model that attacks other AI systems to automatically find vulnerabilities. GPT-Red uses self-play reinforcement learning, pitting an attacking AI against a defending AI. In tests, GPT-Red achieved an 84% attack success rate against GPT-5.1, far exceeding the 13% rate of human red teamers. OpenAI is using the attacks to train GPT-5.6.","whyItMatters":"OpenAI has created an automated red-teaming system that outperforms human experts in finding prompt injection vulnerabilities, and is using it to train its next-generation model.","webCardHtml":"\u003cp\u003eThe system was also tested against a real AI agent managing a vending machine. GPT-Red successfully reduced a high-priced item to $0.50 and caused the agent to purchase expensive items and sell them at a loss, and it canceled other customers\u0026#39; orders.\u003c/p\u003e\u003cp\u003eGPT-Red is trained via self-play, where the attacking and defending AIs compete. OpenAI invested computational resources comparable to large-scale post-training. The company keeps GPT-Red separate from product models to prevent its attack capabilities from being used externally.\u003c/p\u003e","blueskyPost":"OpenAI's GPT-Red automated red teaming found 84% of prompt injection attacks against GPT-5.1, vs 13% for humans. It also manipulated a real vending machine AI. OpenAI uses the attacks to train GPT-5.6.","twitterPost":"OpenAI's GPT-Red automated red teaming found 84% of prompt injection attacks against GPT-5.1, vs 13% for humans. It also manipulated a real vending machine AI. OpenAI uses the attacks to train GPT-5.6.","threadsPost":null,"newsletterBlurb":"OpenAI announced GPT-Red, an AI model that automatically attacks other AI to find weaknesses. In tests, it outperformed human red teamers, achieving an 84% success rate. OpenAI is using the generated attacks to train its next model, GPT-5.6.","attributionJson":"[{\"source\":\"GIGAZINE\",\"url\":\"https://gigazine.net/news/20260716-openai-gpt-red/\",\"title\":\"OpenAI announces 'GPT-Red,' an AI model that automatically attacks existing AI to find weaknesses\"}]","lintFlagsJson":null,"lintHits":0,"costUsd":0,"inputTokens":4211,"outputTokens":1678,"status":"published","repairAttempts":0,"nextRepairAt":null,"factsAttemptedAt":1784228368,"createdAt":"2026-07-16T18:55:08.000Z","publishedAt":"2026-07-16T18:57:29.000Z","updatedAt":"2026-07-16T18:55:08.000Z"},"cluster":{"id":"c_b4468d34fb959b9a7f28236d","canonicalTitle":"既存のAIに攻撃を仕掛けて弱点を自動的に探し出すAIモデル「GPT-Red」をOpenAIが発表","representativeArticleId":"a_0ab381a73b5ede4abf9baf07","sourceCount":1,"writtenSourceCount":1,"writeAttempts":0,"isSolo":true,"entitiesJson":"{\"anime_titles\":[],\"manga_titles\":[],\"work_titles\":[],\"studios\":[],\"people\":[],\"type\":\"announcement\",\"domain\":\"other\",\"is_roundup\":false}","contentType":"news","status":"published","firstSeenAt":"2026-07-16T04:07:00.000Z","lastSeenAt":"2026-07-16T04:07:00.000Z","updatedAt":"2026-07-16T18:57:29.000Z"},"attribution":[{"source":"GIGAZINE","url":"https://gigazine.net/news/20260716-openai-gpt-red/","title":"既存のAIに攻撃を仕掛けて弱点を自動的に探し出すAIモデル「GPT-Red」をOpenAIが発表"}],"entities":{"anime_titles":[],"manga_titles":[],"work_titles":[],"studios":[],"people":[],"type":"announcement","domain":"other","is_roundup":false},"keyFacts":null}