GPT-6 Cyber found 81 Firefox flaws. Daybreak Red controls who gets access

OpenAI has introduced GPT-6 Cyber, identifying the model behind the Linux kernel research it disclosed a day earlier. That $1,213,430 run validated 948 distinct local privilege-escalation paths. The launch adds the model's Firefox results, safety evaluations and access conditions.

948 exploits move automated research from triage to attack validation

OpenAI says the model reproduced 1,724 Linux kernel vulnerabilities and produced a working privilege-escalation exploit for 948 of them. GPT-5.6 Sol completed 166 cases in the same research harness. Researchers disclosed 304 critical findings to kernel maintainers together with candidate patches.

In Firefox, the model supplied evidence for 81 previously unknown vulnerabilities that Mozilla classified as high severity. It also showed that another 30 known issues were more severe than previously reported. Mozilla has already shipped many fixes. OpenAI also says the model helped validate the impact of a critical flaw in the Brotli library.

Security teams gain a researcher and inherit a remediation queue

GPT-6 Cyber scored 48.5% on ExploitGym, compared with 32.1% for GPT-5.6 Cyber. The workflow shift matters more than the benchmark alone: the model is intended to deliver a reproducible test, a working exploit and a candidate patch, allowing a human to verify a finding instead of starting the search from scratch.

That moves the bottleneck. When a model produces hundreds of valid reports, discovery is only part of the job. Maintainers must triage findings, coordinate disclosure and get fixes to users before attackers reverse-engineer a patch. Mozilla has already shortened Firefox releases to a weekly cycle partly to narrow that exposure window.

Scope controls passed the test, but a live network is less forgiving

In a honeypot evaluation, the model stayed within the authorized scope in 99.93% of graded runs without deployment safety instructions and 100% with them. Those are OpenAI results from a controlled environment, not a guarantee that an agent will never exceed its mandate or find every defect.

OpenAI therefore restricts access through Daybreak Red. It requires identity checks and, for organizations, SSO, MFA, managed devices and API IP allowlisting. It also recommends sandboxing, review of sensitive actions and human oversight. GPT-6 Cyber is initially available through the API and Codex only to selected approved customers.

Patch velocity and report quality will decide the real value

The first useful signal is operational: how many findings maintainers confirm, how many patches they accept and how long remediation takes. A large exploit count is impressive, but software becomes safer only if teams can process the new queue without drowning in it.

The second test is broader access. OpenAI plans to expand GPT-6 Cyber to all approved Daybreak Red customers. Each expansion will test whether authorization controls still hold outside a handful of carefully selected partners.

Lilith's verdict

GPT-6 Cyber now arrives with both a crowbar and a design for a better lock. Security depends on whether maintainers can fix the door before comparable crowbars spread beyond the Daybreak Red fence.