{"rewrite":{"id":"r_00f8006b58ce1bdca5781ac1","clusterId":"c_1c14bdb21533d84258991ffe","slug":"openai-develops-gpt-red-to-attack-llms-claims-gpt-5-6-is-most-secure-yet","model":"deepseek-v4-flash","headline":"OpenAI Develops GPT-Red to Attack LLMs, Claims GPT-5.6 Is Most Secure Yet","summary":"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.","whyItMatters":"The development signals a shift toward adversarial training as a core method for hardening large language models against real-world attacks.","webCardHtml":"\u003cp\u003eOpenAI 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.\u003c/p\u003e","blueskyPost":"OpenAI's GPT-Red is an AI designed to break other AIs. The same method that hardened GPT-5.6 could also be repurposed as an offensive tool. OpenAI is both the defender and the one building the weapon.","twitterPost":"GPT-Red is an AI built to attack LLMs. The same technique that made GPT-5.6 secure could also be used to break other models.","threadsPost":"OpenAI built GPT-Red, an AI that attacks other language models. The company used it to harden GPT-5.6, which it now calls its most secure model. The tension: the same adversarial training method that improves security is also a blueprint for offensive AI. OpenAI is both the shield and the sword.","newsletterBlurb":"OpenAI has developed GPT-Red, a super-advanced hacker AI that attacks other large language models. Through repeated training battles, the company says GPT-5.6 became its most secure model to date. The full report is on MIT Technology Review.","attributionJson":"[{\"source\":\"ASCII.jp\",\"url\":\"https://ascii.jp/elem/000/004/420/4420007/?rss\",\"title\":\"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\"}]","lintFlagsJson":null,"lintHits":0,"costUsd":0,"inputTokens":5790,"outputTokens":1350,"status":"published","repairAttempts":0,"nextRepairAt":null,"factsAttemptedAt":1784153534,"createdAt":"2026-07-15T22:04:28.000Z","publishedAt":"2026-07-15T22:07:29.000Z","updatedAt":"2026-07-15T22:04:28.000Z"},"cluster":{"id":"c_1c14bdb21533d84258991ffe","canonicalTitle":"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","representativeArticleId":"a_54e36b9d2d65c3aa67b1d7e7","sourceCount":1,"writtenSourceCount":1,"writeAttempts":0,"isSolo":true,"entitiesJson":"{\"anime_titles\":[],\"manga_titles\":[],\"work_titles\":[],\"studios\":[],\"people\":[],\"type\":\"news\",\"domain\":\"other\",\"is_roundup\":false}","contentType":"news","status":"published","firstSeenAt":"2026-07-15T20:58:11.000Z","lastSeenAt":"2026-07-15T20:58:11.000Z","updatedAt":"2026-07-15T22:07:29.000Z"},"attribution":[{"source":"ASCII.jp","url":"https://ascii.jp/elem/000/004/420/4420007/?rss","title":"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"}],"entities":{"anime_titles":[],"manga_titles":[],"work_titles":[],"studios":[],"people":[],"type":"news","domain":"other","is_roundup":false},"keyFacts":null}