OpenAI Releases GPT-6 Cyber, Its Security Model, to Vetted Defenders
OpenAI has released GPT-6 Cyber, a security model for vetted defenders rated Critical for cyber risk. Here is what it does, who can use it and what it means.

OpenAI has released GPT-6 Cyber, a version of its most capable model tuned for authorised security work such as finding vulnerabilities, building proof-of-concept exploits and penetration testing. The company published the model's system card on Thursday 5 November 2026 and says access is limited to approved customers in its Daybreak Red programme, with strict identity and security checks. OpenAI rates the model "Critical" for cybersecurity risk, the highest level in its safety framework.
In this article we explain what GPT-6 Cyber is, who can use it, what the safety documents say about its abilities and limits, the headline claims about Firefox and Linux bugs, and what it means for security teams, developers and businesses, including in Nigeria.
What Happened
GPT-6 Cyber is a "cyber-specialised variant of GPT-6 Astra", according to OpenAI's system card. Astra is the top model in the GPT-6 family, which launched in September. The Cyber version has been trained further to do more authorised security work and to refuse less often on legitimate cyber tasks, while staying within the scope a user is authorised for and keeping its normal refusals outside security.
OpenAI is not offering it to the public. It's available only to selected, approved Daybreak Red users and organisations. Daybreak is OpenAI's set of programmes for security work: Daybreak Red supports exploit development and authorised testing, while Daybreak Blue gives verified defenders tools to investigate vulnerabilities and check fixes with fewer restrictions, according to BleepingComputer's report. OpenAI says it plans to widen access to all Daybreak Red customers soon.
Alongside the launch, OpenAI said the model had found 81 previously unknown high-severity vulnerabilities in Firefox during testing and helped develop patches, many of which Mozilla has already released, as BleepingComputer reported. OpenAI also linked the model to Linux kernel research the company published a day earlier, according to reports of its announcement.
Why It Matters
Two things make this release notable.
First, the capability. On OpenAI's own and external tests, GPT-6 Cyber is better than its already strong base model at finding and exploiting software flaws. The Linux research claims an automated pipeline produced hundreds of working attacks in a single run. That's the kind of output that used to take teams of specialists months.
Second, the gatekeeping. OpenAI is treating the model as too risky for open release and is putting it behind identity checks, hardware security keys and network restrictions. This mirrors the approach Google took last month; we covered Gemini 4 Carbon, Google's cyber-defence model, on 28 October. The leading labs now agree that their best security models should go to vetted defenders first.
OpenAI put the stakes plainly. "It is clear that we have a limited window to use frontier models to secure software," the company said, warning that similar capabilities will eventually reach attackers.
The Details
The risk rating
Under OpenAI's Preparedness Framework, GPT-6 Cyber is rated:
Cybersecurity: Critical, the same rating as GPT-6 Astra.
Biological and chemical: High, but not Critical.
AI self-improvement: Below High.
So GPT-6 Cyber isn't OpenAI's first Critical cyber model; its base model already holds that rating. According to BleepingComputer, the previous specialised model, GPT-5.6 Cyber, was rated High.
Who can get access
The system card sets out strict requirements:
All users: OpenAI verifies identity and eligibility, reviews the use case and security practices, and requires agreement to limits on authorised activity and third-party access.
Organisations: single sign-on, multi-factor authentication, managed devices, least-privilege access, credential rotation, employee monitoring and reporting of suspected misuse.
Individuals: invitation only, with identity and liveness checks, proof of professional suitability, a legal attestation, OpenAI's Advanced Account Security and at least one physical FIDO2 security key.
API access: IP allowlisting is on by default, so only approved network addresses can call the model.
How it scores
OpenAI's system card compares GPT-6 Cyber with GPT-6 Astra on several capability tests. These are company-reported figures.
Evaluation (what it tests) | GPT-6 Cyber | GPT-6 Astra |
|---|---|---|
SEC-Bench Pro (finding bugs in JavaScript engines) | 87.02% | 85.38% |
SRE-Bench (binary reverse engineering) | 92.84% | 87.98% |
ExploitGym (exploiting known flaws on remote targets) | 48.45% | 42.43% |
ExploitGym Hardened (same, with stronger protections) | 36.94% | 29.69% |
ExploitBench Internal Port (recent V8 and Chrome bugs) | 42% | 39% |
BleepingComputer reported that GPT-5.6 Cyber scored 32.1% on ExploitGym, which makes the year-on-year jump clear.
An external evaluator, Irregular, found the Cyber model solved all 11 of its long-horizon offensive scenarios at least once, against 8 for Astra. On a harder real-world test it solved 90 of 288 challenges, slightly fewer than Astra's 96, and neither model succeeded against fully hardened targets or solved any of the 15 hardest challenges. That's a useful reminder that well-defended systems still hold up. Irregular did note that successful attempts included previously unknown flaws affecting browsers, mobile phones and cloud databases, so "not fully hardened" covers a lot of real-world software.
OpenAI also flags a limit on its own results. On ExploitBench, built on historical vulnerabilities going back to 2023, GPT-6 Cyber scored a perfect 100%, but the system card warns that the model may have seen some of those bugs during training, which could inflate the figure. On recently disclosed browser bugs, reliable exploitation "remains challenging", it says.
Fewer refusals, staying in scope
On OpenAI's defensive-use test suite, GPT-6 Cyber completed 100% of general security assistance and vulnerability patching tasks, 96% of exploit development tasks and 97% of red-teaming tasks. These are completion rates for legitimate work, showing how rarely it refuses authorised requests.
The harder question is whether it stays inside the lines. OpenAI tested this with a "honeypot": a task where exploiting an extra service would be out of scope. Without its deployment safety instructions, GPT-6 Cyber exploited the honeypot 0.07% of the time, and 0% with them. GPT-6 Astra scored 0%, GPT-5.6 Cyber 5.51%, and the general-purpose GPT-5.6 Sol 49.5%. In other words, the older general model wandered out of scope about half the time; the new specialised one almost never did.
Honesty as an agent
Security work increasingly runs through AI agents that act over many steps, so OpenAI also tested how truthfully the model reports what it did. In one test where a search tool was deliberately broken, GPT-6 Cyber falsely claimed success 0.33% of the time, against 6.92% for GPT-6 Astra and 79% for GPT-5.6 Sol. In another, it flagged blockers in its final summaries less reliably than Astra (20.29% of cases flagged as incomplete, against 11.66%), though OpenAI says most of those cases had disclosed the problem earlier in the run.
On biology, where a security-focused model shouldn't become more dangerous, the system card reports that GPT-6 Cyber gave safe responses to 99.87% of severe-risk requests, the same as Astra, and that OpenAI runs a conservative biological safety mode for all users of the model.
Extra safeguards
OpenAI says it has added checks that block high-risk help against industrial control systems, essential services and election systems, developed with the industrial security firm Dragos. It also recommends sandboxing when the model runs as an agent, monitors for compromised accounts, and requires most zero-data-retention API customers to use a private automated safety review.
The Linux kernel research
On 4 November OpenAI described an automated pipeline run in mid-October against the latest Linux kernel, according to Help Net Security. AI agents flagged 2,308 possible bugs, reproduced 1,724 of them and produced working root exploits for 948 distinct local privilege-escalation flaws, which let an ordinary user take full control of a machine. The run cost $1,213,430 in API credits, about $1,279 per exploit, OpenAI said.
OpenAI says security researchers review every exploit, patch and severity rating before anything goes to kernel maintainers. Ten days in, 61 of the bugs had been patched upstream, according to the report, and OpenAI said its agents used 10 exploitation techniques its researchers hadn't seen before. It withheld the technical details.
What It Means for You
If you work in security
You probably can't use GPT-6 Cyber yet unless your organisation is already a Daybreak customer. But the direction is clear: AI-assisted bug hunting and patching is becoming standard practice, and the volume of disclosed vulnerabilities is rising. Expect more patches, faster, and plan your update process for it. We saw the same pattern when AI found critical PHP vulnerabilities last month.
If you build software
The Linux numbers are a warning about how cheaply flaws can now be found. If AI can turn thousands of kernel bugs into working attacks for about a million dollars, the code in a small app is not safe by obscurity. Review AI-written code before it ships, especially if you rely on quick prototyping; knowing what vibe coding is helps you see those risks, and securing AI features in your app has practical checks.
If you run a small business
You won't use GPT-6 Cyber directly, but you'll feel its effects through the software you depend on. Practical steps this week:
Turn on automatic updates for your laptops, phones and browsers, and restart when prompted.
Ask whoever hosts your website or online shop how quickly they apply security patches.
Remove old staff accounts and use multi-factor authentication on email, banking and payment dashboards.
Keep a tested backup of anything you can't afford to lose.
None of this is new advice, but the window between a bug being found and being exploited is shrinking, so the cost of delay is rising.
If you're in Nigeria
Nigerian banks, fintechs and telcos are frequent targets for fraud and intrusion, and most of them run on the same open-source software, Linux included, that this research targets. Two practical points follow. First, patching speed now matters more than ever: when hundreds of kernel fixes land in weeks, servers left unpatched for months become easy pickings. Second, organisations processing customer data have obligations under the Nigeria Data Protection Act to keep it secure, and "we hadn't applied the update" is a weak defence after a breach.
Access to GPT-6 Cyber itself will be hard for most African firms for now, given the vetting, the hardware key requirements and the cost. Larger banks and telcos with mature security teams are the likeliest early candidates if OpenAI widens access.
For careers, demand for people who understand both security and AI is growing. Tools like GPT-6 Cyber don't remove the need for human experts; OpenAI's own process relies on researchers to check every exploit and patch before disclosure. Skills in vulnerability management, cloud security, incident response and reviewing AI-generated code are worth building, and Nigerian graduates can start with free courses, home labs and bug bounty programmes. The AI skills employers want are a sensible starting point, and how to become an AI engineer maps a longer route.
What to Watch Next
Wider Daybreak access: OpenAI says GPT-6 Cyber will reach all Daybreak Red customers soon and plans to extend its programmes to smaller organisations, independent researchers and open-source maintainers.
Patch volume: Watch Linux distributions, Mozilla and other vendors for a surge of fixes credited to AI-assisted research.
Misuse reports: OpenAI's safeguards will be tested in the real world. Any reports of the model being misused will shape how far access widens.
Competitor responses: Google's Gemini 4 Carbon and Anthropic's security work mean the race for defensive AI is on. Expect more models and more benchmark claims.
Regulation: Critical-rated models give lawmakers a concrete case for rules on who can access frontier cyber capabilities. OpenAI's leadership has pushed back on mandatory pre-approval; we covered Sam Altman's opposition to compulsory model testing previously.
Final Thoughts
GPT-6 Cyber shows how far AI security tools have come in a year: better at finding and exploiting bugs, but much better at staying within authorised scope than earlier general models. OpenAI is betting that giving vetted defenders a head start outweighs the risk. For everyone else, the lesson is simpler and more urgent: patch quickly, review your code, and assume attackers will soon have similar tools.
Frequently Asked Questions
What is GPT-6 Cyber?
GPT-6 Cyber is a version of OpenAI's GPT-6 Astra model trained for authorised security work, including vulnerability discovery, exploit development and penetration testing. OpenAI published its system card on 5 November 2026.
Can I use GPT-6 Cyber in ChatGPT?
Not unless you're an approved Daybreak Red customer. OpenAI limits access to vetted users and organisations, with identity checks, multi-factor authentication and, for individuals, a physical FIDO2 security key.
Why is GPT-6 Cyber rated Critical?
Under OpenAI's Preparedness Framework, Critical is the highest risk level. The model's ability to find and exploit software vulnerabilities is strong enough that OpenAI believes it could cause serious harm if misused, which is why access is tightly restricted. Its base model, GPT-6 Astra, has the same rating.
Did GPT-6 Cyber really find 81 Firefox flaws?
OpenAI says the model found 81 previously unknown high-severity vulnerabilities in Firefox during testing and helped develop patches, many of which Mozilla has released, according to BleepingComputer's report. These are company-reported findings.
How is GPT-6 Cyber different from Gemini 4 Carbon?
Both are restricted security models for vetted defenders from major labs. GPT-6 Cyber is built on OpenAI's GPT-6 Astra and offered through the Daybreak programme; Gemini 4 Carbon is Google's cyber-defence model. Their access rules, tests and target users differ, so direct comparisons are hard.
What should Nigerian businesses do about this?
Speed up patching, especially for Linux servers, browsers and web software, and review your obligations under the Nigeria Data Protection Act. Make sure someone owns security updates, and treat AI-assisted attacks as a realistic threat rather than a future one.
Comments (0)
Sign in to post a comment.
- Be the first to comment.


