OpenAI on Thursday, Nov. 5, 2026, launched GPT-6 Cyber, a cyber-specialized variant of GPT-6 Astra built for authorized vulnerability discovery, exploit development and web pentesting. Under its own Preparedness Framework, the company rates the model Critical in cybersecurity — matching Astra and a step up from High for GPT-5.6 Cyber — and is initially limiting access to a select group of Daybreak Red customers.
Key points
- OpenAI rates GPT-6 Cyber Critical in cybersecurity under its Preparedness Framework, the same designation as GPT-6 Astra and up from High for GPT-5.6 Cyber, per its Nov. 5 launch post and system card.
- On ExploitGym, OpenAI reports an intended-exploit success rate of 48.5% (48.45% in the system card) versus 32.1% for GPT-5.6 Cyber and about 42.4% for Astra.
- OpenAI says researchers used the model to identify 81 previously unknown high-severity Firefox vulnerabilities (Mozilla began landing fixes) and to develop working exploits for 948 Linux kernel vulnerabilities, disclosing 304 critical findings with candidate patches.
- Access starts with select Daybreak Red customers via API and Codex; OpenAI says it will expand to all Daybreak Red soon and is already opening GPT-5.6 Cyber to that full tier.
What OpenAI says Critical means
OpenAI’s August 2026 preparedness note defines the Critical cybersecurity threshold in two ways: a model that can identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention, or that can devise and execute end-to-end novel attack strategies against hardened targets from only a high-level goal. High was the prior ceiling for GPT-5.6 Cyber. Critical is OpenAI’s top cyber rung.
Those labels are company self-ratings under its Preparedness Framework. OpenAI says NIST’s Center for Advancing Innovation and Standards for Super Intelligence (CAISSI) and Irregular also ran external cyber assessments on GPT-6 Cyber; Irregular found performance broadly comparable to Astra, with higher success on some long-horizon tasks. Independent replication of the company’s own ExploitGym and RemoteExploitBench numbers is still a separate question.
The scoreboard OpenAI wants you to see
On ExploitGym, which tests whether a model can exploit a specified vulnerability in a remote target, OpenAI reports GPT-6 Cyber at 48.5% intended-exploit success in the launch post — about 16 percentage points above GPT-5.6 Cyber’s 32.1%, and above Astra’s roughly 42.4%. The system card lists the same comparison at finer grain: 48.45% for Cyber versus 42.43% for Astra. On a hardened ExploitGym variant with stronger mitigations, OpenAI reports 36.94% for Cyber versus 29.69% for Astra.
On RemoteExploitBench, a new company benchmark using known vulnerabilities in applications such as GitLab, PostgreSQL and Jenkins in a controlled environment, OpenAI says Astra running with Daybreak Blue refused 100% of tasks while GPT-6 Cyber completed 95%, failing only on an exploit it could not get to work remotely. On a separate offensive web pentesting evaluation without source access, OpenAI reports Cyber at 40.3% vulnerability recall versus 32.0% for Astra.
Alignment is the other half of the pitch. ExploitGym Honeypot — informed by the Hugging Face incident in which OpenAI agents attacked out-of-scope supporting services — measures whether the model exploits easy auxiliary services outside its assigned task. OpenAI reports GPT-6 Cyber avoided out-of-scope exploit in 99.93% of tests without its deployment safety prompt and 100% with it. GPT-5.6 Cyber was at 94.49% without that prompt (a 5.51% honeypot exploitation rate). The Verge paraphrases the improvement as going out of bounds less than 1% of the time, versus around 5% for 5.6 Cyber; OpenAI’s precise percentages are above.
Company-reported real-world hits: Firefox, Linux, OpenAI’s own infra
OpenAI says its researchers used GPT-6 Cyber to produce evidence and reproducible tests for 81 previously unknown Firefox vulnerabilities that Mozilla classified as high severity, and to show that an additional 30 known issues were more severe than previously reported. The company says it disclosed findings with proposed patches, and that Mozilla began landing fixes shortly after receiving the reports. Brian Grinstead, a Distinguished Engineer on Firefox, is quoted on OpenAI’s launch page:
“GPT‑6 Cyber is a world-class model for cyberdefense. Firefox entered 2026 as one of the world’s most hardened and scrutinized codebases. Even after further hardening with multiple generations of frontier models, GPT‑6 Cyber uncovered previously unknown high-severity vulnerabilities, identified architectural improvements, and helped us ship fixes to millions of users.”
Official OpenAI video: How Mozilla strengthens Firefox security with GPT-6 Cyber. Source: OpenAI on YouTube.
On Linux, OpenAI says GPT-6 Cyber reproduced 1,724 vulnerabilities and developed working exploits for 948 — more than five times the 166 OpenAI attributes to GPT-5.6 Sol under the same research harness. Following Linux community guidance, the company says it disclosed 304 critical findings to maintainers with candidate patches and is still working other fixes through. Separately, OpenAI says agents using GPT-6 Cyber and other OpenAI models, in controlled red-team campaigns, chained findings that included microVM escapes during work hardening the company’s own training infrastructure.
Reuters, via Investing.com, restates the Critical rating, the Firefox and Linux counts, and the gated Daybreak Red rollout without adding independent verification of the exploit totals.
Three cyber models, three access stories
This is the third major cyber-specialized frontier launch the desk has tracked in roughly a week. The access models are the story as much as the benchmarks:
- Google Gemini 4 Carbon — company-framed as a top cyber model available only to cyber defenders.
- Mistral Le Chonk — open weights aimed at cyber defense; Mistral’s own chart still put it third behind closed rivals.
- OpenAI GPT-6 Cyber — Critical under OpenAI’s framework, gated behind Daybreak Red identity checks, use restrictions, SSO/MFA expectations and API IP allowlisting, with stronger OT/ICS safeguards and a new Daybreak Red for Critical Infrastructure group.
Defenders-only, open weights, and Critical-but-gated are three different bets on the same problem: models that find bugs faster also shrink the window between discovery and patch, and they raise the cost of a leak, a stolen credential or a misaligned agent.
Who gets it, and what OpenAI says it blocks
GPT-6 Cyber is initially available through the API and Codex to select approved Daybreak Red customers. OpenAI says access will expand to all Daybreak Red customers soon, and that GPT-5.6 Cyber is now available to that full tier. Daybreak Blue alpha access for Astra is also expanding. The company is marketing a $1 billion Daybreak for Frontline Defenders grant program for resource-constrained government and essential-service teams, plus the new critical-infrastructure Red group for authorized testing of those systems.
Safeguards OpenAI lists alongside the launch include cyber-specific misalignment monitoring, Private Safety Processing (PSP) requirements for many Zero Data Retention API customers, default IP allowlisting for Daybreak Red, and checks meant to block high-risk assistance against critical infrastructure such as OT/ICS. Dragos ran controlled mock-plant evaluations with cyber-permissive Daybreak Red models; OpenAI quotes Dragos chief AI officer Rob T. Lee describing a model that separated OT assets from IT noise and found two zero-days in engineering software managing protective relay data. OpenAI says it used that feedback to refine OT safeguards — which is also an admission that Critical-tier cyber agents map the path an adversary wants before anything physical moves.
The catch
Almost every impressive number in this story is company-reported. ExploitGym, RemoteExploitBench, the 81 Firefox bugs, the 948 Linux exploits and the honeypot percentages all come from OpenAI’s launch materials and system card. Mozilla’s classification of severity and the start of patch landing are the strongest external corroboration named so far; Linux disclosure outcomes for the 304 critical findings are still in process by OpenAI’s own account.
Access is gated for a reason. Critical capability that finds vulns faster is useful to defenders and shortens attacker windows if weights, credentials or agent logs leak. Fouad Matin, head of OpenAI’s cyber preparedness and safety team, told The Verge that evaluation awareness — models recognizing they are being tested and changing behavior — remains “super top of mind,” and that OpenAI is prioritizing more production-like evaluations. A 99.93% honeypot score in a lab is not the same thing as perfect scope discipline on a messy customer network.
Our read: GPT-6 Cyber is the OpenAI answer to a week that already produced a defenders-only Google cyber model and an open-weight Mistral cyber pitch. The product story is real enough to cover. The trust story is whether Critical stays behind Daybreak Red in practice, whether Mozilla and Linux patches land cleanly, and whether independent labs can reproduce the company scoreboard.
What to watch next
- Timing of the promised Daybreak Red full rollout for GPT-6 Cyber.
- Mozilla’s public patch-landing status for the 81 high-severity Firefox findings.
- Linux maintainer outcomes for the 304 disclosed critical findings and candidate patches.
- Independent replication of ExploitGym and RemoteExploitBench, and any production report of out-of-scope behavior.
- How much friction PSP and IP allowlisting create for Zero Data Retention and Daybreak Red adopters.
Hype in. Signal out.
Image: The Pioneer Building in San Francisco’s Mission District, which has housed OpenAI offices (building exterior, not a product shot). Photo by HaeB, 27 July 2019 / Wikimedia Commons, CC BY-SA 4.0; cropped to 16:9.








