OpenAI launched GPT-6 Cyber on November 5, a GPT-6 Astra variant specialized in authorized security research. Initial API and Codex access is limited to selected, approved Daybreak Red customers. The official announcement includes Firefox and Linux findings that illustrate the goal: turning suspected flaws into evidence people can verify and fix.
What the tests show
OpenAI reports 48.5% for GPT-6 Cyber on ExploitGym, compared with 32.1% for GPT-5.6 Cyber. That is a 16.4 percentage point difference in the vendor's own evaluation. This score does not measure the probability of compromising an arbitrary real system.
The company also attributes 81 previously unknown Firefox vulnerabilities, rated high severity by Mozilla, to the model. Some of this work has public corroboration: the Firefox 159 security advisory, published on October 27, credits multiple findings to OpenAI's Amy Burnett. It documents specific fixes, without establishing that all 81 cases have been resolved.
Linux: reproducible results still need review
The Linux kernel campaign report, published on November 4, describes 1,724 reproduced vulnerabilities and 948 flaws with working local privilege escalation exploits. That does not mean remote attacks against 948 servers.
The experiment consumed around $1.21 million in API credits. This is the reported run cost, rather than a complete operating budget. Human researchers reviewed the results before reporting them to maintainers. A proposed patch is also distinct from a fix that has been merged and shipped.
For software maintainers, those distinctions matter. A suspected flaw starts an investigation; a reliable reproduction supports prioritization; a validated fix reduces exposure. Counting everything as “bugs found” hides the work between these stages.
What the Critical designation means
The GPT-6 Cyber system card assigns it Critical cybersecurity capability, matching Astra under the Preparedness Framework. This is a model capability category, rather than an incident alert.
Access requires verification and organizational controls. The document also acknowledges limitations: alignment results have strengths and weaknesses relative to Astra, including worse performance at reporting blockers in the final response. Staying within scope remains something to verify during execution.
What changes for technology professionals
If you build or operate systems, the potential benefit is better evidence for deciding what to fix first. A useful evaluation for your team should track reproducibility, review time and patch quality alongside alert volume.
That also changes what you ask an agent to deliver. Beyond a list of problems, request reproducible evidence, the conditions needed to trigger a flaw and a way to test the fix, always within an authorized environment.
Finding more vulnerabilities helps defenders when their process can absorb the findings. Is your team ready to turn that volume into patched software?