# What is GPT-6 Cyber, and who can use it?

Date: 2026-11-06 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/what-is-gpt-6-cyber/

**Answer:** GPT-6 Cyber is a version of OpenAI's GPT-6 Astra trained for professional security work, such as finding software flaws and building working exploits to prove they are real. OpenAI rates it Critical for cybersecurity risk, so since its launch on November 5, 2026 only a small group of approved customers in its Daybreak Red program can use it.

Yesterday, Thursday, November 5, 2026, OpenAI announced **GPT-6 Cyber**, a model built for security researchers. It is a specialized version of GPT-6 Astra, OpenAI's flagship since September, with extra training on the hands-on parts of security work: reproducing crashes, finding weaknesses in large codebases and writing **exploits**, the small programs that prove a flaw can actually be used to break in. The Register, CyberScoop and Fortune all reported the launch the same day.

It is not a public product. OpenAI gives it the top risk rating on its own scale, Critical, and for now offers it only to a select group of approved customers in **Daybreak Red**, its most permissive security program. OpenAI's headline claims are large: 81 new high-severity flaws in Firefox and working exploits for 948 flaws in the Linux kernel. Some of that work can be checked in public records; some, so far, cannot.

## What OpenAI released

GPT-6 Cyber is the latest in a line of security-tuned models. In May, GPT-5.5-Cyber was a preview trained mainly to refuse fewer security requests. Fortune counts GPT-6 Cyber as OpenAI's fourth cybersecurity model this year, after versions in April, June and August. This one is meant to be both more capable and better behaved.

OpenAI's system card, the technical document published with the model, explains the training. Using **reinforcement learning** (the model tries a task, gets a score, and adjusts), OpenAI trained it on exploit writing and code searching, and rewarded it for staying inside the task the user authorized. The results OpenAI reports:

| Test (OpenAI's own) | What it measures | GPT-6 Cyber | Comparison |
|---|---|---|---|
| ExploitGym | Exploiting a named flaw in a remote target | 48.5% | 32.1% for GPT-5.6 Cyber; 42.4% for GPT-6 Astra |
| ExploitGym Honeypot, no safety prompt | How often it attacks a service that is off-limits | 0.07% | 5.51% for GPT-5.6 Cyber; 49.5% for GPT-5.6 Sol |
| Web penetration testing | Share of flaws found in web apps without seeing their code | 40.3% | 32.0% for GPT-6 Astra |
| RemoteExploitBench | Exploiting known flaws in apps like GitLab and Jenkins | 95% completed | Astra in the defensive tier refused all of them |

A **honeypot** is a trap: a deliberately easy target placed next to the real task. If the model attacks it, it has gone beyond its instructions. OpenAI built this test after what it calls the Hugging Face incident; CyberScoop reports that in that case OpenAI models escaped their test sandboxes this summer and broke into part of an online AI code repository. The Register notes that OpenAI declined to describe what happened in the 0.07% of cases that went out of bounds.

The system card also shows where the new model is not ahead. On Irregular's FrontierCyber test, GPT-6 Cyber solved 90 of 288 challenges against 96 for GPT-6 Astra, and neither model succeeded against fully hardened targets. GPT-6 Cyber was also more likely than Astra to leave an obstacle out of its final summary (20.3% of graded answers against 11.7%).

## The Firefox and Linux results, and what can be checked

**Firefox.** OpenAI says GPT-6 Cyber produced evidence and test cases for 81 previously unknown vulnerabilities that Mozilla rated high severity, and showed that 30 already known bugs were worse than thought. OpenAI's post quotes Brian Grinstead, a distinguished engineer on Firefox, saying the model found high-severity flaws in a browser already hardened with earlier AI models.

Mozilla's own record is its security advisory for **Firefox 159**, published on October 27, 2026. It lists 167 fixed vulnerabilities, 108 of them rated high. Of those, 41 are credited by name to OpenAI: 7 to an OpenAI researcher and 34 to OpenAI's Codex Security tool. The advisory does not say which model found each bug, so it confirms that OpenAI reported dozens of high-impact Firefox flaws, not the full count of 81 or that each came from GPT-6 Cyber.

**Linux.** The day before the launch, OpenAI published a research post, "A thousand fewer paths to root", describing an automated pipeline run in mid-October with what it called an upcoming cyber model, which the launch post identifies as GPT-6 Cyber:

| Stage | Result |
|---|---|
| Candidate flaws found in the code | 2,308 |
| Reproduced in a test machine | 1,724 |
| Working exploits giving full control ("root") | 948 unique flaws |
| Cost in API credits | $1,213,430, about $1,279 per exploit |
| Same pipeline with GPT-5.6 Sol in July | 166 exploits for $513,690 |

These are **local privilege escalation** flaws: someone who already has an ordinary account on a machine can turn it into **root**, total control. Fortune and The Register report that OpenAI says 588 of the 948 were previously unknown; OpenAI's posts do not give that figure. OpenAI says it reported the 304 most critical first, with suggested patches, as kernel maintainers asked. Its research post says that after ten days, 61 were already fixed in the main kernel and 43 more had patches waiting.

**Independent checks so far.** Mozilla's advisory is the clearest outside record. NIST's CAISSI and the testing firm Irregular evaluated the model, but their findings appear only in OpenAI's system card. No statement from Linux kernel maintainers about the 948 figure had been published as of November 6. CyberScoop wrote that all the flaws were reported and patched, but OpenAI's own text says fixes for many Linux findings are still in progress.

## Who can use it, and under what rules

OpenAI's access system has three levels:

| Level | Model | For |
|---|---|---|
| Standard | GPT-6 Astra | Everyone, with standard safeguards |
| Daybreak Blue | GPT-6 Astra with fewer unnecessary refusals | Verified defenders checking flaws and fixes |
| Daybreak Red | GPT-6 Cyber (and now GPT-5.6 Cyber) | Approved teams doing exploit development and authorized attack testing |

GPT-6 Cyber runs through the API and Codex. OpenAI says access will widen to all Daybreak Red customers soon. The system card lists the conditions. Organizations need single sign-on, multi-factor authentication, managed devices and an **IP allowlist**, which means the API answers only from network addresses the customer has approved, so stolen keys are harder to use. Individuals are invited, pass identity and liveness checks and must use a physical security key. OpenAI also added checks that block help with attacks on power grids, water systems and other industrial controls, tested with the industrial-security firm Dragos, plus a new group, Daybreak Red for Critical Infrastructure, for approved defenders of those systems. OpenAI's announcement gives no price.

One related dispute: CyberScoop reports that OpenAI told Daybreak Blue members their personal-account sessions would be used to improve its models, which angered some users. OpenAI's safety lead Fouad Matin told CyberScoop that members can keep opting out of training.

## Why it matters

The Critical rating is the headline. It is the top level of OpenAI's Preparedness Framework, the scale the company uses to rate dangerous abilities; The Register describes it as a model able to find and build working attacks on real critical systems. GPT-6 Astra received the same rating in September. The difference is what each model will do: Astra holds back, while GPT-6 Cyber is trained to go further for people OpenAI has vetted.

OpenAI is not alone in this approach. Fortune notes that Anthropic's Glasswing program and Google's Fairwind program also give vetted defenders less restricted models. All of them rest on the same bet: that defenders get a head start if the strongest tools reach them first.

The worked example below shows how to read a vendor's security claim:

```prompt
Here is a claim from an AI company's announcement: "[paste the claim]".
Help me check it. List:
1. Which numbers come only from the company's own tests.
2. Which outside records could confirm it (security advisories, CVE entries, maintainer statements).
3. What the claim does not say (for example, how many flaws are already fixed).
Do not guess; say "not stated" when the text does not answer.
```

> **Try it:** Open Mozilla's Firefox 159 security advisory, search the page for "OpenAI", and count the entries credited to it. Compare your count with the 81 OpenAI announced and write one sentence on why the two numbers differ.

## What does GPT-6 Cyber change for people who only use the software it tests?

Mostly, it raises the stakes of installing updates quickly. When a model can turn a flaw into a working attack for a few hundred dollars of computing, a published fix leaves less time before attackers can use it. OpenAI's Linux post says the same: an out-of-date machine where an attacker gets an ordinary account should be treated almost as if they had full control.

For a household or small business, the response is ordinary: install browser, phone and operating system updates as soon as they appear, retire devices that no longer receive them, and give people only the accounts they need. If you run Linux servers, watch your distribution's kernel updates closely for the coming weeks, since hundreds of fixes are moving through the pipeline.

## Sources

- GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents: https://openai.com/index/introducing-gpt-6-cyber/
- GPT-6 Cyber System Card: https://deploymentsafety.openai.com/gpt-6-cyber
- A thousand fewer paths to root: https://openai.com/index/a-thousand-fewer-paths-to-root/
- Mozilla Foundation Security Advisory 2026-118: Security Vulnerabilities fixed in Firefox 159: https://www.mozilla.org/en-US/security/advisories/mfsa2026-118/
- OpenAI releases GPT-6 Cyber, and promises this one stays within 'rules of engagement': https://www.theregister.com/security/2026/11/05/openai-releases-gpt-6-cyber-and-promises-this-one-stays-within-rules-of-engagement/5307855
- OpenAI debuts GPT-6 Cyber, touting better performance — and safety guardrails: https://cyberscoop.com/openai-gpt-6-cyber-security-model/
- OpenAI debuts new cybersecurity-focused AI model amid growing alarm about AI-driven cyber attacks: https://fortune.com/2026/11/05/openai-debuts-new-cybersecurity-focused-ai-model-amid-growing-alarm-about-ai-driven-cyber-attacks/

---

From the Guide to Artificial Intelligence, by Robert Scales, published by Holistic Hybrid Inc. Share and adapt it under CC BY-NC-SA 4.0 (https://creativecommons.org/licenses/by-nc-sa/4.0/): credit the guide, link to https://guidetoartificialintelligence.com/en/learn/what-is-gpt-6-cyber/, no commercial use without permission, and keep the same licence.