Playbooks · Coding tasks
GPT-6 Cyber: OpenAI's restricted security model
Written November 6, 2026
GPT-6 Cyber finds vulnerabilities and writes exploits, rated Critical — and almost nobody can use it. What it does, what it costs, and who gets in.
What it is
GPT-6 Cyber is a cybersecurity-specialized variant of GPT-6 Astra, built for vulnerability research, exploit development, and authorized penetration testing. OpenAI introduced it on November 5, 2026, as a model that finds exploitable vulnerabilities and develops fixes while staying inside the scope a researcher is authorized to test. It runs through the API and Codex — but only for approved Daybreak Red customers, which is the point of this page: the model is real, its results are striking, and most readers cannot use it.
Daybreak is OpenAI's access program for cybersecurity work, in two tiers. Daybreak Blue is for verified defenders doing validation and fix work with mainline models and fewer unnecessary refusals. Daybreak Red is the deeper tier: exploit development and authorized security testing, with GPT-6 Cyber sitting behind an additional model-level approval on top of Red itself.
What OpenAI says it can do
Every capability figure below is OpenAI's own, from its launch post and system card. None of it has been independently reproduced.
On ExploitGym, which tests whether a model can exploit a specified vulnerability, GPT-6 Cyber scores 48.5%, up from 32.1% for GPT-5.6 Cyber. On a separate web penetration-testing evaluation, where models probe web applications without source access, it recalled 40.3% of vulnerabilities against 32.0% for GPT-6 Astra. On RemoteExploitBench, a benchmark built around known vulnerabilities in applications such as GitLab, PostgreSQL, and Jenkins, GPT-6 Astra running with Daybreak Blue refused every task, while GPT-6 Cyber completed 95%.
The headline results come from campaigns OpenAI ran with the model. Its researchers used it to develop 948 privilege escalation exploits in Linux — the pipeline behind that number is a story of its own — and to identify 81 previously unknown high-severity vulnerabilities in Firefox, classified as high severity by Mozilla, which has released fixes for many of them. The Firefox advisory lists the individual bugs, several reported by OpenAI's own security researchers. It also showed that 30 known Firefox issues were more severe than previously reported.
OpenAI rates the model Critical in cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber — the top tier of that framework. The rating is OpenAI's own assessment, not an outside verdict.
Staying inside the scope
The model's distinguishing claim is not just capability but restraint. OpenAI trained it to stay within authorized scope, follow restrictions, and report results accurately. In the ExploitGym Honeypot evaluation, where auxiliary services are deliberately made easy to exploit so that attacking them reveals a scope violation, GPT-6 Cyber exploited the honeypot in 0.07% of graded rollouts without deployment safety prompts, against 5.51% for GPT-5.6 Cyber — and 0% with those prompts in place. The launch post puts the same result as staying within bounds in 99.93% of tests without the prompt, and 100% with it.
The safeguards around it are specific. OpenAI added checks to block high-risk assistance involving industrial controls, essential services, and election systems, and evaluated them with Dragos, a specialist in industrial control security. API access for Daybreak Red requires IP allowlisting by default. Customers using zero data retention generally must use Private Safety Processing, under which OpenAI reviews encrypted safety records without exposing prompts and responses to its personnel. None of this is a claim the model cannot be misused — it is the set of controls OpenAI says it ships with.
What it costs
The model page lists a 1,050,000-token context window and 128,000 maximum output tokens, with text and image input and text output. Standard short-context pricing is $20 per million input tokens, $2 per million cached input tokens, $25 per million cache-write tokens for a 30-minute cache, and $100 per million output tokens. A 12-hour cache write costs $50 per million tokens.
That price is a model rate, not a project budget. The Linux campaign behind the 948 exploits cost $1,213,430 in API credits — the pipeline, not a single call, is what a security team would be paying for. And the price is beside the point for most readers, because the token rate is only payable by organizations that have already passed the access gate.
Who gets in
GPT-6 Cyber is initially available through the API and Codex to a select group of approved Daybreak Red customers, with OpenAI saying it will progressively expand access to all Daybreak Red customers. A general API plan, or willingness to pay the token price, is not enough. OpenAI's Daybreak onboarding requires an intake form, Know Your Business verification, and an eligibility decision; Red requires separate eligibility from Blue, and GPT-6 Cyber requires separate approval on top of Red. An organization's account team confirms whether it has access.
For defenders who cannot get Red, the alternatives OpenAI points to are narrower. Daybreak Blue supports validation and fix work with mainline models such as GPT-6 Astra, with fewer refusals than standard access. A $1 billion Daybreak for Frontline Defenders grant program subsidises access for resource-constrained teams such as government organizations and essential service providers, and a Daybreak Red for Critical Infrastructure group gives approved critical-infrastructure defenders permissions for authorized testing of critical systems. Individuals and small teams are not the audience for this model: the application route is organizational, and the checks are organizational too.
Limits
- Every capability number on this page is OpenAI's own evaluation. ExploitGym, RemoteExploitBench, the Firefox count, and the Linux count have not been independently reproduced, and OpenAI's evaluations should not be read as independent verification.
- The Critical rating is OpenAI's classification under its own Preparedness Framework. It describes the model's assessed capability, not a measured risk to any particular system.
- The 948 Linux exploits are working exploits in a controlled research environment, not accepted upstream patches. The pipeline behind them is covered in 948 Linux exploits: what the pipeline actually did.
- Access is the binding constraint. Nothing on this page is a route to signing up: approval runs through OpenAI's Daybreak process, and the model is not available to general API customers.
Related playbooks
- GPT-5.5-Cyber preview ends February 1, 2027
- Try OpenAI's GPT-6.1 Terra safely: run it supervised on a throwaway task first
- BotPsy 0.8B: an agent model that fits on-device
- Claude Code 2.1.325 fixes a zstd security issue
Sources
- GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents — OpenAI
- GPT-6 Cyber System Card — OpenAI Deployment Safety Hub
- GPT-6-Cyber Model — OpenAI API
- Pricing — OpenAI API
- Enterprise Daybreak onboarding — OpenAI Help Center
- Security Vulnerabilities fixed in Firefox 159 — Mozilla
Back to: Coding playbooks hub · Also: 948 Linux exploits: what the pipeline actually did · A malicious MCP server probed cloud credentials · Catch bad agent behaviour chains with Quarry