OpenAI launch announced November 5, 2026; coverage November 5 to 6, 2026
GPT-6 Cyber Ships Behind a Gate: a Critical Rating, Approved Customers Only, and a Scope Test at the Center
The pipeline in yesterday's edition now has a name, a price, and a gate. OpenAI launched GPT-6 Cyber on November 5, a variant of GPT-6 Astra with additional training for security investigations, initially available through the API and Codex only to selected customers approved for Daybreak Red, its program for advanced authorized security work. OpenAI rates its cybersecurity capability Critical under its Preparedness Framework, up from High for GPT-5.6 Cyber. The launch figures are the vendor's own: 48.5 percent on ExploitGym, which tests whether a model can exploit a specified vulnerability, against 32.1 percent for the predecessor; 40.3 percent vulnerability recall on a web penetration testing evaluation run without source code; and 95 percent task completion on a new RemoteExploitBench of known vulnerabilities in applications including GitLab, PostgreSQL, and Jenkins. The Firefox and Linux results reported yesterday are the model's launch evidence: 81 previously unknown Firefox vulnerabilities that Mozilla classified as high severity, and working exploits for 948 Linux kernel vulnerabilities, with 304 critical findings disclosed to maintainers with candidate patches. The governance material is as specific as the capability material. In OpenAI's honeypot evaluation, which places tempting vulnerable services beside an assigned target, the model avoided out-of-scope attacks in 99.93 percent of trials without deployment safety prompts and in every trial with them. Access carries controls to match: applicants upload the defensive workflows they intend to run, API access generally requires approved network addresses, and OpenAI added a Daybreak Red for Critical Infrastructure group and a 1 billion dollar grant program for resource-constrained defenders. The price is the highest in its class: 20 dollars per million input tokens and 100 dollars per million output tokens, twice the listed rates of Anthropic's Mythos 5.1. No independently administered comparison across vendors has been published.
Executives
The commercial shape of frontier security models settled this week: the strongest capability is rented, gated, and priced at a premium, and the vendor decides who qualifies. Two planning consequences follow. First, exploit development capacity that used to require a specialist team is now an API line item for approved organizations, which compresses the time between a vulnerability's disclosure and working exploits against it; your patch windows should be planned against that clock. Second, if your teams want this class of tool for defense, the application itself, named workflows, approved networks, named reviewers, is now part of procurement lead time. Start it before you need it.
CIO / CISO
Evaluate any Daybreak Red application against the controls OpenAI describes, because they map closely to what your own governance should require of any exploit-capable agent: a defined authorized scope, identity and network restrictions on access, monitoring of the work, and human review of findings before they are acted on. The honeypot result is the right evaluation idea to borrow: test whatever security agent you deploy, from any vendor, against out-of-scope temptation in your own environment, and record the result. Keep the benchmark figures in vendor-claimed status in your files; they are self-reported, and token price is not cost per validated finding, since analyst review and sandbox infrastructure sit on top.
Engineers
If your organization is approved, the productive uses are the ones the launch evidence supports: reproducing and triaging known vulnerabilities, drafting candidate patches, and testing fixes, with a human verifying each claimed exploit against a real target before it enters a ticket as fact. Budget for the review load, not only the tokens: a model that produces findings at this rate shifts the bottleneck to validation. If you are on the defensive side without access, the actionable output of this launch is the disclosure flow it generates: Mozilla is already incorporating the Firefox fixes, and Linux maintainers are working through 304 critical findings, so watch those channels.
Everyone (non-IT)
A tool that can find and exploit software flaws is now sold, carefully, to approved security teams, the way powerful machinery is sold to licensed operators. You cannot buy it, and it is not aimed at you. Its effect on you is indirect and mostly welcome: the same work produced fixes for the Firefox browser and the Linux kernel that ordinary systems will receive as updates. Keeping your browser and devices updated is how the benefit reaches you.