OpenAI opened GPT-6 Cyber to a narrow slice of the security world on November 5, and the part most teams will care about is not the benchmark score. It is the gate. The model is live in the API and inside Codex, but only for approved Daybreak Red customers. Everyone else gets GPT-5.6 Cyber, or waits for Daybreak Blue alpha with GPT-6 Astra.
Here is how to get approved.
Start with the numbers, because they explain the access policy. GPT-6 Cyber scores 48.5% on ExploitGym, the benchmark that asks an agent to turn a known vulnerability into a working exploit, up from 32.1% for GPT-5.6 Cyber. On RemoteExploitBench, which plants known flaws in GitLab, PostgreSQL and Jenkins, it completed 95% of tasks while GPT-6 Astra with Daybreak Blue refused 100% of them. On scope discipline it avoided attacking out-of-scope services in 99.93% of tests without a deployment prompt, and 100% with one. OpenAI now rates the model Critical in cybersecurity under its Preparedness Framework, up from High.
The receipts are grim. OpenAI's own researchers used it to reproduce 1,724 Linux kernel vulnerabilities and build working exploits for 948 of them, more than five times the 166 that GPT-5.6 Sol managed with the same harness. It also produced evidence for 81 previously unknown high-severity Firefox bugs, plus a critical Brotli flaw tracked as CVE-2026-111497 with a CVSS score of 9.2.
- API model ID:
gpt-6-cyber, served through Responses, Chat Completions and Batch endpoints. - Limits: 1,050,000 token context window, 128,000 max output tokens, reasoning effort set to highest.
- Price per million tokens: $20 input, $2 cached input, $100 output, with cache writes at $25 for 30 minutes or $50 for 12 hours.
Getting approved: the practical checklist
Access is an account decision, not a signup form.
- Confirm you qualify as a defender. Current users include Cisco, Calif, CodeRabbit, CrowdStrike, SentinelOne, Palo Alto Networks and Mozilla. Independent researchers and open-source maintainers are being added through Daybreak Blue alpha, which runs GPT-6 Astra rather than Cyber.
- Apply for Daybreak, then ask your account team for Daybreak Red. The docs are explicit: your organization needs Daybreak Red access and approval, and the account team confirms it. Approval applies to the org and the project, so name both in the request.
- Expect IP allowlisting. Daybreak Red API access requires approved network ranges by default. Collect your egress IPs before the approval call, not after.
- Critical infrastructure defenders get a separate lane. Ask about Daybreak Red for Critical Infrastructure, which permits authorized testing of OT and ICS systems while keeping alignment safeguards on.
- Small teams and public bodies should ask about subsidy. The $1 billion Daybreak for Frontline Defenders program covers government organizations and essential service providers that cannot absorb enterprise pricing.
Once approved, keep the first project small and auditable: one repository, one scoped target, one reviewer. The model's value is in the loop of find, validate, patch, verify, and that loop only works if the scope is written down before the agent starts.
- Do not budget for ChatGPT access. GPT-6 Cyber ships to API and Codex customers only.
- Daybreak Blue session content from personal accounts will be used to improve models starting November 29, 2026, even when the improve-the-model setting is off.
- Zero Data Retention API customers generally must use Private Safety Processing, which keeps encrypted records in customer storage for 30 days for automated review.
- Sandbox the agent, review its actions, and keep a human on the high-risk steps. OpenAI recommends all three.
The uncomfortable framing OpenAI puts on its own launch is that the window to use frontier models for defense is limited. That makes the approval queue the product feature that matters most this quarter. If your team owns software that other people depend on, start the Daybreak application today and treat the benchmark charts as the reason the paperwork exists.
Comments