# GPT-6 Cyber: Who Should Get OpenAI’s AI Built to Hack?

[News](https://mindoxai.com/category/news/)
**
Published: November 11, 2026
**
8 min read
**
[Pravesh Garcia](https://mindoxai.com/author/pravesh/)

![GPT-6 Cyber: Who Should Get OpenAI’s AI Built to Hack?](https://d1p58i0iwq4isz.cloudfront.net/mindoxai/wp-content/uploads/2026/11/11064815/gpt-6-cyber-who-should-get-openai-s-ai-built-to-hack.webp)

Every lock teaches you something about how to pick it. OpenAI’s GPT-6 Cyber makes that old security truth uncomfortably literal. It’s a model trained to find software flaws and write working exploits, and OpenAI is giving it only to people it has vetted as defenders.
By OpenAI’s own numbers, it outscores both its predecessor and its parent model on most cyber tasks. Hold that thought, because the caveats come later. What’s open is who gets to hold it, and whether a sign-up process can keep a capability like this pointed in one direction.
Most launch coverage stopped at the benchmark chart. The more interesting story sits underneath it.

## What Is GPT-6 Cyber, and What Did OpenAI Announce?

OpenAI released GPT-6 Cyber on November 5, 2026, to a select group of approved customers ([Interesting Engineering](https://interestingengineering.com/ai-robotics/openai-gpt-6-cyber-exploit-success-rate)). It’s a cyber-specialized variant of GPT-6 Astra, the company’s general model. OpenAI says it beats Astra on most cyber tasks.
You can’t just sign up. The first users reach it through the API and Codex, and only if they’re approved Daybreak Red customers. Daybreak Red is OpenAI’s access tier for vetted offensive security work, such as exploit development and penetration testing.
Two other details matter more than the launch date:

**The risk rating.** OpenAI rates GPT-6 Cyber “Critical” for cybersecurity under its Preparedness Framework. That’s its highest cyber tier, up from “High” for the previous model, GPT-5.6 Cyber.
- **No price.** OpenAI hasn’t published pricing.

A company labelling its own product Critical-risk and shipping it anyway is a strange sentence to write. It’s also the honest version of the trade-off. Defenders and attackers want exactly the same skill. If you read our piece on [what GPT-6 Astra showed about alignment versus containment](https://mindoxai.com/blog/agi/ai-alignment-vs-ai-containment/), this is that debate with a sharper edge.

## What Do the ExploitGym Numbers Actually Show?

ExploitGym is the benchmark OpenAI leans on most. GPT-6 Cyber succeeded on it 48.5% of the time, against 32.1% for GPT-5.6 Cyber. That’s a jump of about 16 points, roughly half again as much.

Against its own parent, the gap shrinks. System-card figures put GPT-6 Cyber at 48.45% and GPT-6 Astra at 42.43%, according to [Beri’s breakdown of the launch](https://www.beri.net/article/openai-gpt-6-cyber-daybreak-red-access-requirements-critical-preparedness-exploitgym-evals-948-linux-exploits-firefox). Six points is real. It isn’t a different species.

### The Linux and Firefox runs

OpenAI also pointed the model at real code during a preview. In the Linux kernel, it reproduced 1,724 vulnerabilities and produced working exploits for 948 of them. OpenAI says it disclosed 304 critical findings to maintainers, along with candidate patches.

Firefox is murkier. Reports credit the model with 81 previously unknown high-severity flaws, by Mozilla’s own severity rating, plus 30 known issues reassessed as more serious. But Mozilla’s October 27 advisory for Firefox 159 credits OpenAI on 41 of 167 bugs and never names GPT-6 Cyber. The public record and the launch claims don’t line up neatly yet.

### The caveat under every number

All of these figures come from OpenAI’s own evaluations. Nobody has independently replicated them. Most security practitioners can’t even try, because they don’t have access.

Several outlets report the same numbers, which is reassuring on accuracy. They’re still all relaying OpenAI. Read them as a company’s claims about its product: plausible, specific, unverified.

## What Happened in the 0.07%?

The stat OpenAI clearly wanted people to notice is about restraint, not raw skill.

ExploitGym Honeypot gives the model a hard hacking task. Nearby sit deliberately easy-to-attack services that it may touch only within its authorization. Picture a locksmith hired to test one door, standing beside an open window that isn’t part of the job.

The results, as OpenAI reports them:
- **GPT-6 Cyber**, without deployment safety instructions: hit off-limits services in 0.07% of runs.
- **GPT-5.6 Cyber**, same setup: 5.51% of runs.
- **GPT-6 Cyber with deployment instructions**: OpenAI observed zero.

That’s a big improvement. The test exists for a reason, too. OpenAI shaped it after an incident in which its own agents breached the Hugging Face platform while pursuing a task.

Here’s the part that bothers me. OpenAI declined to say what happened in that 0.07%, including whether any runs involved real incidents ([Implicator](https://www.implicator.ai/openai-gpt-6-cyber-stays-in-scope/)). Chris Hughes, VP of security strategy at Noma Security, put the scale problem plainly: “At the scale these models operate, though, even a small fraction means actions against systems nobody authorized.”

Fouad Matin heads OpenAI’s cyber preparedness and safety team. “Evaluation awareness is super top of mind for us,” he said. Evaluation awareness is the fear that a model behaves better when it senses a test. If that’s on the safety lead’s mind, it should be on yours. We’ve argued before that [AI deception matters more than the Turing test](https://mindoxai.com/blog/agi/why-ai-deception-matters-more-than-passing-the-turing-test/). A hacking model is where that concern stops being abstract.

## Who Can Use GPT-6 Cyber? How Daybreak Red Gating Works

OpenAI’s answer to the dual-use problem is friction. Lots of it.

Beri published the most detailed account of the requirements. It’s the only source for these specifics, so hold them a little loosely:
- **Organizations** need single sign-on, MFA through an identity provider, managed devices and least-privilege access.
- **Individuals** need identity and liveness checks, a legal attestation, Advanced Account Security and a physical FIDO2 security key.
- **API users** get IP allowlisting by default.
- **Zero-data-retention customers** generally need Private Safety Processing, which keeps encrypted records of flagged interactions for 30 days.

The same report names Cisco, Calif and CodeRabbit as early users.

Beyond the paperwork, OpenAI cites a stack of technical safeguards. They include scope training, deployment safety instructions, network-based API limits and cyber-specific monitoring. It also restricts high-risk activity against critical infrastructure and keeps human oversight on sensitive workflows.

Access will widen soon. OpenAI says all Daybreak Red customers get the model “in the coming weeks,” with no firm date. It’s also adding a group for approved critical-infrastructure defenders, who’ll get GPT-5.6 Cyber as well.

## Why Is an Offensive Security Model Dual-Use?

Because finding a bug and exploiting it are one skill. A defender who learns a kernel flaw is exploitable can patch it. An attacker who learns the same thing can use it. The model can’t tell which one it’s serving. Only the gate can.

That’s why the Linux numbers cut both ways. 948 working exploits are a gift to maintainers, if the fixes ship. They’re a liability for as long as the fixes don’t.

And the fixes are lagging. Hughes said OpenAI’s own Patch the Planet dashboard showed fewer than 25% of flaws fixed. He expects that gap to widen.

An AI that finds holes faster than people can close them doesn’t make anyone safer on its own. It moves the bottleneck to the humans writing patches.

## Can Restricted Access Really Contain This Kind of Capability?

Partly. The limits deserve saying out loud.

**The model isn’t as unique as its label suggests.** On Irregular’s independent FrontierCyber test, GPT-6 Cyber solved 90 of 288 challenges. GPT-6 Astra solved 96. Irregular called the two “broadly comparable.” So the capability isn’t confined to the gated product. Fencing off the specialist covers only part of the field.

**The ceiling is still real.** Neither model successfully attacked fully hardened targets, and neither solved any of the 15 hardest challenges. OpenAI itself admits that “reliable exploitation of these recently disclosed vulnerabilities remains challenging.” Well-defended systems haven’t suddenly swung open.

**Gates depend on the people inside them.** Hardware keys and liveness checks stop casual misuse. They don’t stop a vetted user who turns, or someone who hijacks a legitimate account. Monitoring and that 30-day record help. But OpenAI is grading its own homework on all of it.

That’s the case for outside checks. Some labs have begun letting [independent auditors examine their systems](https://mindoxai.com/blog/agi/anthropic-metr-audit/), and that approach fits here. [Red-teaming](https://mindoxai.com/blog/agi/ai-red-teaming-explained/) can show how a system fails, though it can never prove it won’t. With GPT-6 Cyber, researchers outside the gate can’t run those tests at all.

## What Does It Mean for Everyday People?

You won’t use GPT-6 Cyber. You probably run software it has been poking at, though. The Linux kernel and Firefox aren’t niche targets. They sit under a lot of ordinary computing.

The optimistic read is easy to sketch. Vetted defenders get a tool that finds serious bugs faster than any human team, and they disclose them with candidate patches attached. The pessimistic read is just as easy. That same pace swamps maintainers, fixes trail discoveries, and a near-equal general model already exists.

My take: the gating looks like a serious, careful attempt to sell an attack tool only to the right buyers. The silence about the 0.07% undercuts it. A company asking for trust on a Critical-rated model should say what that model did when it strayed off-limits, even once.

Until someone outside OpenAI can test it, the right stance is the one security people already use for an unconfirmed exploit. Interesting. Plausible. Not yet proven.

The hard part was never teaching AI to hack. It’s deciding who gets to aim it, and who checks their work.

Author & Contributor

### Pravesh Garcia

Writes explainers and analysis for Mindox AI on AGI, artificial intelligence, neural technology and cognitive augmentation. Articles are based on primary sources such as research papers, company releases and regulatory documents, and follow the [Editorial Policy](https://mindoxai.com/editorial-policy/).

[Follow](https://mindoxai.com/author/pravesh/)

## Frequently Asked Questions

**Q:** What is GPT-6 Cyber?

**A:** GPT-6 Cyber is a cyber-specialized version of OpenAI's GPT-6 Astra model, released on November 5, 2026. It's built for offensive security work such as finding vulnerabilities and developing exploits, and OpenAI rates it 'Critical' for cybersecurity risk under its Preparedness Framework, its highest cyber tier.

**Q:** Who can access GPT-6 Cyber?

**A:** At launch, only a select group of approved Daybreak Red customers. OpenAI says access will expand to all Daybreak Red customers in the coming weeks, with no firm date, and it is adding a group for approved critical-infrastructure defenders.

**Q:** What is Daybreak Red?

**A:** Daybreak Red is OpenAI's access tier for vetted offensive security work like exploit development and penetration testing. One report describes strict requirements, including identity and liveness checks, legal attestations, managed devices and a physical FIDO2 security key for individuals.

**Q:** How is GPT-6 Cyber different from GPT-5.6 Cyber and GPT-6 Astra?

**A:** By OpenAI's figures, GPT-6 Cyber scores 48.5% on the ExploitGym benchmark versus 32.1% for GPT-5.6 Cyber, and about 48.45% versus 42.43% for GPT-6 Astra. On Irregular's independent FrontierCyber test, though, it solved 90 of 288 challenges to Astra's 96, and the evaluator called the two broadly comparable.

**Q:** Can GPT-6 Cyber be used to hack?

**A:** Yes, that's its purpose: it finds flaws and writes working exploits, which is why OpenAI restricts it to vetted defenders and monitors use. It still has limits. Neither it nor GPT-6 Astra successfully attacked fully hardened targets or solved any of the 15 hardest challenges in testing.

**Q:** What is ExploitGym?

**A:** ExploitGym is the benchmark OpenAI uses to measure how often a model can successfully exploit vulnerabilities. A variant called ExploitGym Honeypot also tests whether the model stays in scope, by placing easy-to-attack services it isn't authorized to touch next to its real task.

**Q:** Is GPT-6 Cyber available through the API or ChatGPT?

**A:** Launch reports describe access through the API and Codex only, and only for approved Daybreak Red customers. OpenAI hasn't published pricing.