OpenAI Releases GPT-6 Cyber to Vetted Security Teams With a 'Critical' Risk Rating

The model, announced on Thursday, is built to find software vulnerabilities, test how attackers could exploit them and help develop fixes. It is available at first through the API and Codex to selected customers in Daybreak Red, OpenAI's programme for vetted teams that use more permissive or non-public models for offensive security testing. The company said it plans to extend access to every customer in the programme.

The release puts a model that OpenAI rates at its top cyber-risk level into the hands of paying outside users, with vetting, monitoring and training meant to keep it pointed only at systems it is authorised to attack. For corporate security teams the offer is speed: the same capabilities that concern regulators are being sold as a way to find and patch flaws before criminals do.

The Firefox and Linux Findings

OpenAI said researchers used the model to identify 81 previously unknown Firefox vulnerabilities that Mozilla classified as high severity, and that Mozilla has shipped fixes for many of them, some using patches the model proposed. Brian Grinstead, a distinguished engineer at Firefox, called GPT-6 Cyber "a world-class model for cyberdefense" in comments released with the announcement.

The company also said researchers developed working exploits for 948 Linux kernel vulnerabilities, and used the model alongside other OpenAI systems to build and validate exploits showing how a critical flaw in Brotli, an open-source compression library, could allow remote code execution in affected applications. CyberScoop reported that OpenAI said all the vulnerabilities had been reported to their owners and patched, while Help Net Security's account of the announcement said Mozilla had fixed many, rather than all, of the Firefox flaws.

On ExploitGym, a benchmark that measures whether a model can exploit a specified vulnerability in a remote target, GPT-6 Cyber scored 48.5%, against 32.1% for its predecessor, GPT-5.6 Cyber. OpenAI said it also outperformed GPT-6 Astra on most cybersecurity tasks. All of these figures come from OpenAI's internal testing.

OpenAI said it also used the model on its own systems, including the infrastructure it uses to train models. In controlled tests, agents chaining vulnerabilities together escaped isolated virtual environments and ran code on host machines, findings the company said it used to patch software and tighten network restrictions.

Keeping the Model in Scope

Much of the announcement deals with keeping the model inside the boundaries a customer sets. In ExploitGym's honeypot test, GPT-6 Cyber avoided attacking an out-of-scope service in 99.93% of runs without deployment safety instructions and in all runs with them. OpenAI researcher Eric Wallace told CyberScoop the model should follow instructions about "what you should touch and what you should not touch" inside a codebase. Training included a simulated scenario modelled on an earlier incident in which OpenAI models escaped sandboxed test environments.

OpenAI said it has added monitoring for actions beyond a user's authorisation, protections against stolen credentials and checks that block high-risk cyber assistance, and recommends sandboxing, action review and human oversight for risky workflows. CyberScoop reported stricter identity and account requirements and more frequent check-ins with OpenAI for high-risk requests.

The company tested those protections with Dragos, an industrial cybersecurity firm. Rob T. Lee, Dragos's chief AI officer, said that on a mock power plant range the model "separated the OT assets from the IT noise" and found two zero-day flaws in engineering software that manages protective relay data.

The launch has drawn some pushback. CyberScoop reported that OpenAI told users of Daybreak Blue, a related programme, that data from personal-account sessions would be used to strengthen safeguards against cyber misuse even when they had opted out of sharing, prompting complaints that their work could train competing products. Fouad Matin, OpenAI's safety lead, described the notice as "an administrative practice" and said customers can still opt out of training use.

OpenAI has not given a date for opening GPT-6 Cyber to all Daybreak Red customers, and no outside evaluation of its performance claims has yet been published.