# OpenAI's New Model Found 81 Unknown Firefox Flaws. Almost No One Is Allowed to Use It. > GPT-6 Cyber is OpenAI's most capable security model yet, built specifically for the kind of vulnerability hunting that can either protect software or weaponize it. That's exactly why OpenAI is keeping access tightly restricted. - Source: Morning Week - Canonical URL: https://morningweek.com/article/openai-s-new-model-found-81-unknown-firefox-flaws-almost-no-one-is-allowed-to-use-it - Author: Morning Week Newsroom - Section: Tech - Published: 2026-11-05T18:15:39.063Z - Updated: 2026-11-05T18:15:39.063Z - Tags: OpenAI, cybersecurity, AI models, GPT-6 Cyber, vulnerability research, AI safety --- OpenAI has released a new AI model built specifically for cybersecurity work, and the results from internal testing help explain why the company is being unusually cautious about who gets to use it. GPT-6 Cyber, the company's latest specialized model, outperforms its general-purpose counterpart, GPT-6 Astra, on most cybersecurity tasks, with a particular edge in advanced vulnerability research and exploit development. ## What the model actually found The scale of what GPT-6 Cyber uncovered during testing is striking on its own. The model identified 81 previously unknown, high-severity vulnerabilities in Firefox, flaws that had gone undetected despite the browser's heavy scrutiny from security researchers worldwide. It went further with the Linux kernel, developing working exploits for 948 separate vulnerabilities, including 588 that were entirely novel. That output represents more than five times the number of Linux exploits produced by GPT-5.6 Sol, the previous state-of-the-art model in this space, in testing earlier this year. ## A deliberately narrow rollout Capabilities like that cut both ways, and OpenAI appears to know it. Rather than releasing GPT-6 Cyber broadly, the company is launching it through a restricted access program called Daybreak, initially limited to vetted red team customers, security researchers and organizations OpenAI has approved for that level of access. The company has signaled it plans to expand availability gradually, but the cautious rollout reflects a straightforward concern: a model this good at finding and exploiting vulnerabilities could do just as much damage in the wrong hands as it could good in the right ones. ## The fourth release in a fast-moving lineup GPT-6 Cyber is OpenAI's fourth specialized cybersecurity model released this year alone, a pace that reflects both rapid internal progress and mounting external pressure. The company brought on Dali Rajic as its new chief revenue officer, drawing on his background in cybersecurity software to help steer how these increasingly powerful tools get positioned commercially. ## An industry racing to keep up with itself OpenAI is not alone in trying to walk this line. Anthropic has built a comparable vetted-access framework called Glasswing, and Google has rolled out its own version known as the Fairwind Program, both aimed at giving security professionals access to frontier AI capabilities without handing the same tools to attackers. The urgency behind all three efforts is backed by the broader numbers: Gartner has forecast that worldwide spending on AI-powered cybersecurity tools will reach 51.3 billion dollars in 2026, nearly double the 25.9 billion dollars spent the year before. The same capability that lets a model patch a vulnerability before an attacker finds it is the capability that could let an attacker find it first, and that tension now sits at the center of how every major AI lab is deciding what to release and to whom. For now, GPT-6 Cyber's most dramatic results, the kind of vulnerability discovery that once took skilled human researchers months, are staying behind a tightly controlled door. Whether that restraint holds as competitive pressure builds, or as rival labs push their own cyber models further, may say as much about the future of software security as anything the model itself is capable of finding. --- Originally published by Morning Week. Free to cite with attribution and a link to https://morningweek.com/article/openai-s-new-model-found-81-unknown-firefox-flaws-almost-no-one-is-allowed-to-use-it.