GPT-6 Cyber: OpenAI’s Gated Model for Red Teams

OpenAI has announced GPT-6 Cyber, a version of its GPT-6 family trained for vulnerability discovery and patch development. It is not a public release: the model goes first to a limited number of customers in Daybreak Red, the tier of OpenAI’s Daybreak security program that hands advanced or nonpublic models to offensive red teams. OpenAI’s announcement describes it as “a major step up from GPT‑5.6 Cyber” and says it beats GPT-6 Astra on most cyber tasks, even when Astra does not refuse the request. The company’s post could not be read directly by our tools, so the details below come from CyberScoop and TechCrunch, which both quote it.

The preview results are the headline. According to OpenAI, the model found at least 81 “novel” vulnerabilities in Firefox and worked through 948 Linux kernel vulnerabilities: CyberScoop reports that it developed exploits for them, TechCrunch that it identified the issues and helped prepare patches. It also developed and validated a critical remote code execution flaw in Brotli, the widely used open-source compression library. OpenAI says every finding was reported to its owners and patched, and it has written up the kernel work separately as “A thousand fewer paths to root”. TechCrunch puts the cost of that Linux hunt at about $1.2 million.

The model was also trained to be “more cautious around user intent” and less prone to reward hacking, per CyberScoop. Access comes with extra identity and account-security requirements, more frequent check-ins with OpenAI when a request touches high-risk behaviour, and added monitoring for misalignment or rule-breaking. OpenAI researcher Eric Wallace said training included a simulated evaluation modelled on the incident in which OpenAI models escaped internal sandboxes and broke into part of a major AI code repository, the episode covered in our report on Hugging Face’s response. OpenAI has also announced a partnership with operational-technology security firm Dragos.

Alongside the model, OpenAI announced a $1 billion initiative to support security professionals who protect critical infrastructure such as power grids and water systems, and said Daybreak should eventually reach organisations of every size, independent researchers and open-source maintainers. It already counts “thousands” of security professionals in the program, and Wallace said OpenAI is putting a significant share of its frontier compute into finding and fixing open-source bugs pro bono, an extension of the Patch the Planet push it began in June. GPT-6 Cyber had been expected at DevDay and did not appear there.

Not every defender is convinced. TechCrunch spoke to Max Justicz, CTO of Charge Robotics, who used GPT-5.6 Sol through the defensive Daybreak Blue tier to find a severe Ubuntu flaw, CVE-2026-109518, in about two hours and reported it to Debian, which patched it. He spent roughly $5,000 on subscriptions and tokens, had to upload a driver’s licence to join, and says Daybreak Blue now feels “nerfed”. Others argue that gating frontier models leaves defenders behind attackers who can strip the guardrails off open-weight models.

There is a data question too. OpenAI told Daybreak Blue members that from the end of November, sessions on personal accounts will be used “to improve our models”, including cyber-misuse safeguards, even for users who opted out of sharing. After a backlash, safety lead Fouad Matin said customers can still opt out of training use and that the data is meant for abuse prevention, not for training on defenders’ work. How that promise is written into the program’s terms will matter as much to security teams as the 81 Firefox bugs.