OpenAI’s GPT-6 Cyber goes to vetted defenders first, not to the public API

OpenAI has added GPT-6-Cyber to its API models page. The page puts it in a separate “Specialized models” group under the Daybreak cyber program rather than among the flagship GPT-6 models. Access is limited to approved customers.

Listed under Daybreak, with no price or model ID

The entry appeared on OpenAI’s API models page in a section called “OpenAI Daybreak”, described as advanced cyber models for defenders. The page describes GPT-6-Cyber as a model for three jobs: investigating a vulnerability, reproducing it in an authorized test environment, and verifying a fix. Beside it sits GPT-5.6 Cyber, which the page describes as a cybersecurity model for authorized vulnerability research and security testing.

The listing does not show what the flagship cards do: no price per million tokens, no context window, no knowledge cutoff and no reasoning levels. By comparison, GPT-6 Astra is listed at $10 per million input tokens and $50 per million output tokens, with a 1.05M-token context window. Whether GPT-6-Cyber is billed like Astra is not stated on the page.

OpenAI announced the model on X on November 5. It said GPT-6 Cyber pairs stronger capabilities with better model alignment, that it powered the Linux kernel research announced a day earlier, and that security teams at companies were already using it. The post is truncated in the copy we could read.

Who gets access, and when

According to CSO Online, GPT-6 Cyber is available through the API and Codex to a select group of approved Daybreak Red customers. Daybreak is OpenAI’s trusted-access cyber program. It has two tiers: Red, which is more permissive, and Blue, which CSO says is recommended for most defenders. OpenAI plans to extend access to all Daybreak Red customers over the coming weeks, per the same report. GPT-5.6 Cyber is also being offered to that customer set, and an alpha of Daybreak Blue for GPT-6 Astra is being widened.

For most developers this means the model cannot simply be called with an API key. Approval into the program comes first. CSO Online does not report pricing, and neither does the models page.

The research behind it

On November 4, OpenAI said its security researchers had built an autonomous pipeline for finding, exploiting and patching vulnerabilities. In a single run with an upcoming cyber model, it validated and generated patches for 948 local privilege escalation vulnerabilities in the Linux kernel. Reporting by Neowin says 61 had been patched upstream, 43 had staging patches and 299 were under public discussion. CSO Online says 588 of the 948 were previously unknown. It also reports that OpenAI found 81 previously unknown high-severity vulnerabilities in Firefox, with Mozilla beginning to ship fixes.

Outlets tied the kernel work to GPT-6 Cyber. TechRadar noted that OpenAI’s own report refers only to an upcoming cyber model, so the link between the two comes from the November 5 post and press coverage rather than from the original write-up.

A Critical risk rating and a thin remediation record

The safety framing is the main story here. CSO Online reports that OpenAI classifies GPT-6 Cyber as “Critical” for cybersecurity under its Preparedness Framework. That is why the rollout is gated. OpenAI also cites an internal ExploitGym Honeypot evaluation in which the model strayed out of scope in 0.07% of cases, against 5.51% for GPT-5.6 Cyber. These are OpenAI’s numbers, and no outside evaluation has been published.

The same report raises the counterweight: finding bugs is not the same as fixing them. A Noma Security researcher cited by CSO pointed to OpenAI’s Patch the Planet dashboard, which shows a remediation rate below 25%. OpenAI gave no data on how much its models speed up patch deployment at enterprises. Defenders may get faster discovery, but a larger pile of confirmed vulnerabilities only helps if maintainers and IT teams can work through it.

OpenAI also says that, in a sample of 166 exploits, it found 10 techniques that looked novel to outside kernel experts. That is the dual-use problem in concrete form: the same capability that helps a defender reproduce a flaw also lowers the cost of writing an exploit. The tiered Red and Blue structure is OpenAI’s answer, trading breadth of access for vetting.

What the sources leave open

Several things are not established. The page does not give a price, context window or rate limits, and the model ID shown in the title, gpt-6-cyber, comes from the listing’s name rather than a spec card. Public availability is not promised; the only timeline is wider access for Red customers. The Firefox and kernel figures come from OpenAI and press coverage of its report, with technical details withheld. The company’s safety numbers have not been independently checked, and the page was published without a date.