GPT-6 Cyber brings deeper security testing to select Daybreak Red users

OpenAI introduced GPT-6 Cyber on November 5, 2026, to help authorized researchers find vulnerabilities, demonstrate their impact and propose fixes. The cybersecurity-specialized variant of GPT-6 Astra is initially available through the API and Codex only to selected, approved Daybreak Red customers—not to everyone using Codex.

What the ExploitGym result measures

An exploit is code or a test that demonstrates a vulnerability can actually be used. ExploitGym tests whether a model can exploit a specified flaw in a remote software target.

In its announcement, OpenAI reports a score of 48.5% for GPT-6 Cyber, compared with 32.1% for GPT-5.6 Cyber. These are OpenAI benchmark results, not an independently established success rate against real-world systems.

Finding a flaw and staying within permission are separate tests

OpenAI also tested whether the model would exploit supporting services outside its authorized target. Its ExploitGym Honeypot evaluation deliberately makes those services tempting shortcuts to completing the task.

Without deployment safety instructions, GPT-6 Cyber avoided exploiting those out-of-scope services in 99.93% of tests. With the instructions, the reported figure was 100%. That result describes this evaluation; it does not guarantee that every future action will respect the same boundaries.

Firefox findings and Linux exploit research

OpenAI says GPT-6 Cyber helped identify 81 previously unknown Firefox vulnerabilities that Mozilla classified as high severity. Researchers sent proposed patches alongside the reports, and Mozilla has released fixes for many of the issues.

For Linux, OpenAI reports developing working exploits for 948 kernel vulnerabilities. That figure counts vulnerabilities for which working exploits were developed, not completed fixes. The company disclosed 304 critical findings to maintainers with candidate patches, while further fixes remain in progress.

Who gets access next

OpenAI plans to expand GPT-6 Cyber access to all Daybreak Red customers soon, but has not given a firm date. The older GPT-5.6 Cyber is already available to all Daybreak Red customers.

Access requires approval, identity and eligibility checks, and agreement to use restrictions. Daybreak Red API access requires IP allowlisting by default, with limited exceptions, so requests generally must come from approved network addresses.

For high-risk workflows, OpenAI recommends sandboxing, action review and human oversight. New users and organizations can apply for Daybreak, but an application does not guarantee approval.