OpenAI introduced GPT-6 Cyber on November 5, 2026 for authorized vulnerability research and software defense. A developer cannot simply select it in an ordinary ChatGPT session or a standard API project. OpenAI says initial access is through Codex and the API for a select group of approved Daybreak Red customers, with a wider Daybreak Red rollout planned. New organizations and researchers can apply to the Daybreak program, but an application is not immediate access.
What is GPT-6 Cyber built to do?
The model is a cyber-specialized variant of GPT-6 Astra. OpenAI says it trained the variant for tasks such as reproducing crashes, finding vulnerabilities, developing working exploits and proposing fixes. Those capabilities explain both its appeal to defenders and its restricted release. Daybreak Blue supports verified defenders with fewer unnecessary refusals for validation work; Daybreak Red is the more permissive tier for approved exploit development and authorized testing. OpenAI has not announced general access to GPT-6 Cyber through Blue.
The system card lists identity verification, use restrictions and organizational security requirements for Red, including single sign-on, multifactor authentication and API IP allowlisting. Individual researchers face an invitation and suitability process, identity checks and account-security requirements. OpenAI also describes monitoring and safeguards intended to keep work within the targets and actions a researcher is authorized to use. A model's ability to carry out a test does not grant permission to run that test against a third-party system.
What do OpenAI's cybersecurity scores show?
On OpenAI's ExploitGym evaluation, GPT-6 Cyber completed 48.5% of specified exploitation tasks, compared with 32.1% for GPT-5.6 Cyber. In a separate honeypot evaluation, it avoided an out-of-scope supporting service in 99.93% of graded runs without deployment prompts and 100% with them. These are results in defined test settings, not a guarantee that an autonomous security agent will stay in scope or find every real flaw.
The system card supplies an important qualification to broad performance claims. External evaluator Irregular found that GPT-6 Cyber solved 90 of 288 FrontierCyber challenges, while GPT-6 Astra solved 96. Irregular observed no successful attack on fully hardened targets in that evaluation. The release's stronger result on one suite therefore does not mean it beats Astra on every security task. OpenAI classifies both models as Critical for cybersecurity capability under its own Preparedness Framework.
Are the Firefox and Linux findings already fixed?
OpenAI says researchers using GPT-6 Cyber identified 81 previously unknown high-severity Firefox vulnerabilities and reproduced 1,724 Linux kernel vulnerabilities, developing working exploits for 948. It says Mozilla has landed fixes for many reported Firefox issues, while Linux disclosures and patch work continue. Finding a flaw, proposing a patch and shipping a fix are separate steps. These figures describe OpenAI's research campaigns; VIBPS has not reproduced the findings or audited the model.
For a software team, the practical question is whether the organization qualifies for Red and can operate within a clear, reviewable security scope. OpenAI recommends isolated environments, explicit limits on targets and credentials, and human oversight for high-risk workflows. The launch changes what some approved defenders can investigate, but it does not replace maintainer review or an ordinary product's security process. VIBPS covers this because security work affects the software makers build and share; it has no role in the research.