GPT-6 Cyber: what it is and who can use it
2026-11-06 · 4 min read
GPT-6 Cyber is OpenAI's security research model, a version of GPT-6 Astra trained for authorized work like finding vulnerabilities, building exploits and testing fixes. Most people can't use it yet: it's limited to select, approved Daybreak Red customers through the API and Codex.
The posts on X
@OpenAI announced it on Thursday, November 5, calling it "a major advance in security research." The post says the model powered the Linux kernel research OpenAI announced a day earlier, and it points to security teams at companies including Mozilla, Palo Alto Networks, Calif and Cisco. OpenAI president Greg Brockman (@gdb) said it's "already being used to help defend security-critical software."
A second post from @OpenAI sets the goal of Daybreak access for "organizations of every size, independent researchers, and open-source maintainers." The model follows GPT-5.6-Cyber, which we covered in August.
What it can do
These scores come from OpenAI's system card, except the GPT-5.6 Cyber ExploitGym figure, which is from The Register.
| Test | GPT-6 Cyber | Compared with |
|---|---|---|
| ExploitGym | 48.45% | GPT-6 Astra 42.43%, GPT-5.6 Cyber 32.1% |
| ExploitGym Hardened | 36.94% | GPT-6 Astra 29.69% |
| Irregular FrontierCyber | 90 of 288 solved | GPT-6 Astra 96 of 288 |
| Exploited an out-of-scope honeypot, no safety instructions | 0.07% | GPT-6 Astra 0%, GPT-5.6 Cyber 5.51%, GPT-5.6 Sol 49.5% |
| Same honeypot test, with safety instructions | 0% | Not listed |
Lower is better in the honeypot rows, because that test counts how often the model attacked something it wasn't allowed to touch.
OpenAI's researchers also used the model to find 81 previously unknown high-severity Firefox vulnerabilities, and The Register says Mozilla has started fixing them.
The Linux number is bigger. @OpenAI described an autonomous pipeline that validated and generated patches for 948 local privilege escalation vulnerabilities in a single run. The Hacker News puts the bill at $1,213,430 in API credits, or $1,279 per exploit, and The Register says 588 of the bugs were previously unknown. Security researchers review each exploit and patch before it goes to maintainers.
How to get access
Access goes through an application. OpenAI's second post links to a Daybreak sign-up page, and there are two tiers. As Vaibhav Srivastav (@reach_vb) puts it, Blue helps you find and fix bugs with GPT-6 Sol and Luna, while Red is for testing exploits with GPT-5.6 Cyber. Only select Red teams get GPT-6 Cyber today, and The Register reports that all Red customers should get it over the coming weeks.
The system card lists what Red requires:
- Individuals need an invitation, identity and liveness checks, proof of professional suitability, a legal attestation, Advanced Account Security and at least one physical FIDO2 key.
- Organizations need SSO, MFA, managed devices and least-privilege access.
- API access uses IP allowlisting by default.
We found no price in the system card or the coverage we read. If you're looking at Blue instead, check the data terms: The Hacker News reports that starting November 29, OpenAI plans to use content from Blue sessions on personal accounts to improve its models, even with the "improve the model for everyone" setting off.
Our read
The scope numbers interest us more than the exploit count. CyberScoop reports that training included a simulated evaluation meant to mimic the Hugging Face incident, and the drop from 5.51% to 0.07% suggests that work paid off. But OpenAI declined to tell The Register what happened in that 0.07%, and nearly all of these figures come from OpenAI itself.
The system card has less flattering results too. The model lost to Astra on FrontierCyber, evaluator Irregular saw no successful attacks on fully hardened targets, and neither model solved any of the 15 hardest challenges. GPT-6 Cyber also does worse than Astra at reporting blockers in its final answer (flagged in 20.29% of graded responses against 11.66%), though OpenAI says that in most flagged samples it reviewed, the model had mentioned the blocker earlier.
Anthropic is working the same ground: it recently detailed five exploit chains built by an unreleased Claude model.
What it means for businesses
You'll probably never touch this model, but remember the $1,279. That's the API cost of one working route to root in OpenAI's run, and The Hacker News describes that run as taking about 30 hours. When hundreds of flaws can be worked through that fast, the gap between a fix shipping and you installing it matters more. Help Net Security says Mozilla has released fixes for many of the Firefox bugs, so update Firefox and keep your Linux servers on a regular patch schedule.
The honeypot row applies if you're giving your own agents access to company systems. For this model, written limits took out-of-scope attacks from 0.07% to 0% in testing, so spell out what each agent may touch and check that it's enforced. New Face Design's free process audit can map what your AI tools reach today.