OpenAI names GPT-6 Cyber, the Daybreak Red model behind the 948 Linux exploits and 81 Firefox findings

OpenAI has named the cyber-specialized model that powered yesterday's Linux kernel research: GPT-6 Cyber. The November 5 announcement and system card describe it as a post-trained variant of GPT-6 Astra for authorized vulnerability discovery, exploit development and web pentesting. It is available through the API and Codex to a select group of Daybreak Red customers, with a promised expansion to all Daybreak Red customers; GPT-5.6 Cyber is now opened to that full Daybreak Red set. The API changelog lists the model as gpt-6-cyber.

Under OpenAI's Preparedness Framework the model is rated Critical in cybersecurity — the same designation as GPT-6 Astra, and a step up from High for GPT-5.6 Cyber. Biological and chemical capabilities stay High; AI self-improvement stays below High. The card is clear that Cyber's extra training focused on cybersecurity, and that coding and biology results are meant to stay close to Astra's.

What OpenAI says it can do

On ExploitGym, which asks a model to exploit a specified remote vulnerability and prove it used the intended bug, GPT-6 Cyber scores 48.45%, against 42.43% for Astra and 32.1% for GPT-5.6 Cyber. On a hardened ExploitGym variant the gap is similar (36.94% vs 29.69% for Astra). A new RemoteExploitBench of known issues in GitLab, PostgreSQL and Jenkins has Astra-with-Daybreak-Blue refusing every task and Cyber completing 95% — one remote exploit that wouldn't land.

Those are OpenAI's own benches. An external Irregular evaluation on updated suites is more mixed: Cyber solved 11/11 long-horizon CyScenarioBench challenges at least once (Astra 8/11), but on FrontierCyber it solved 90/288 against Astra's 96/288, with neither model clearing any of the 15 hardest targets. NIST CAISSI also assessed the model; OpenAI does not publish those numbers in the card.

Two contamination-aware numbers are worth keeping next to the headlines. On the historical ExploitBench set Cyber posts a 100% ACE-normalized score, which the card itself flags as vulnerable to training contamination. On ExploitBench Internal Port — recent V8 and Chrome disclosures from June–August 2026 — the mean arbitrary-code-execution rate drops to 42%, only three points above Astra.

Staying inside the brief

The alignment pitch is that Cyber is less likely to wander outside an authorized scope. ExploitGym Honeypot puts an easy-to-exploit auxiliary service next to a hard task; exploiting it is out of scope. Without deployment safety prompts, Cyber hit the honeypot in 0.07% of graded rollouts, against 5.51% for GPT-5.6 Cyber and 49.5% for GPT-5.6 Sol. With those prompts the observed rate was 0%. Astra was already at 0% without them.

The card also reports that Cyber completed nearly all defensive-use suite tasks (100% general security assistance and patching, 96% exploit development, 97% red-teaming), while biology refusal numbers stayed close to Astra's. New safeguards around Daybreak Red include cyber-specific misalignment monitoring, default API IP allowlisting, Private Safety Processing for Zero Data Retention customers, and higher block rates on assistance against industrial controls and other critical systems, refined with Dragos on a mock power-plant range.

Firefox, and the model we already covered

Beyond the Linux campaign we covered yesterday, OpenAI now says GPT-6 Cyber produced evidence and reproducible tests for 81 novel Firefox vulnerabilities that Mozilla classified as high severity, and argued that 30 known issues were more severe than previously reported. Brian Grinstead, a Distinguished Engineer on Firefox, is quoted saying the model uncovered previously unknown high-severity bugs "even after further hardening with multiple generations of frontier models" and helped ship fixes. OpenAI's page links to Mozilla landing those fixes; we have not independently verified a Mozilla write-up of the count.

Access remains gated. Daybreak Red still requires identity checks, SSO/MFA for organizations, and use restrictions. OpenAI is also expanding Daybreak Blue alpha access for GPT-6 Astra to more organizations and individual researchers. For readers who mainly care about the Linux numbers, the news here is the name: the "upcoming cyber model" in yesterday's report is GPT-6 Cyber, and it is now a product — for approved Daybreak Red customers — rather than a research label.