OpenAI launched GPT-6 Cyber on 5 November 2026, a cybersecurity model built to find and prove exploitable software flaws and help write the fixes. OpenAI says it has already produced evidence for 81 previously unknown high-severity Firefox vulnerabilities, and Mozilla's fixes sit in Firefox 159. Access is limited to approved security customers, so for most businesses the real change is a faster stream of patches that someone has to install.
What happened
In its announcement of 5 November 2026, published on its own site, OpenAI introduced GPT-6 Cyber as a model trained to find and validate exploitable vulnerabilities and develop fixes, while staying inside the scope a security researcher sets. A vulnerability is a weakness in software; an exploit is working code that uses it to break in.
OpenAI rates the model Critical in cybersecurity under its Preparedness Framework, the internal scale it uses to judge a model's capability and risk. Its previous security model, GPT-5.6 Cyber, was rated High. OpenAI also says GPT-6 Cyber outperforms its general model, GPT-6 Astra, on most cyber tasks.
The announcement leans on real software rather than test scores alone. OpenAI says GPT-6 Cyber produced evidence and reproducible tests for 81 novel vulnerabilities that Mozilla classified as high severity, and showed that another 30 known issues were more serious than first reported. The findings went to Mozilla with proposed patches, and OpenAI says Mozilla began landing fixes shortly after.
Mozilla's own record fits that account. Mozilla Foundation Security Advisory 2026-118, published on mozilla.org and announced on 27 October 2026, covers the vulnerabilities fixed in Firefox 159 and is rated high impact. It lists 167 separate CVE entries (CVE is the public numbering system for known flaws), and seven of them name Amy Burnett of OpenAI as the reporter. Mozilla now issues an entry for every individual bug, which partly explains the long list.
The model was also behind the Linux kernel work OpenAI published a day earlier, which we covered in our post on the 948 Linux kernel exploits found by AI agents. The announcement confirms that GPT-6 Cyber reproduced 1,724 kernel vulnerabilities, built working exploits for 948 of them, and that OpenAI sent 304 critical findings to the maintainers with candidate patches.
Key details
| Item | What OpenAI or Mozilla says |
|---|---|
| Announced | 5 November 2026, by OpenAI |
| Purpose | Find and validate exploitable vulnerabilities and develop fixes, within the scope researchers set |
| Risk rating | Critical in cybersecurity under OpenAI's Preparedness Framework (GPT-5.6 Cyber: High) |
| ExploitGym score | 48.5%, up from 32.1% for GPT-5.6 Cyber, a 51% relative improvement |
| Web testing without source code | 40.3% recall of vulnerabilities, against 32.0% for GPT-6 Astra |
| Staying in scope | Avoided out-of-scope targets in 99.93% of honeypot tests without safety prompts (GPT-5.6 Cyber: 94.49%), and 100% with them |
| Firefox | 81 novel high-severity vulnerabilities; fixes in Firefox 159 (advisory 2026-118, 27 October 2026) |
| Who can use it | A select group of approved Daybreak Red customers, through the API and Codex, expanding to all Daybreak Red customers |
ExploitGym is a test of whether a model can exploit a named weakness in a remote target. The honeypot test is the one OpenAI stresses most: it plants easy, tempting systems next to the real task and checks whether the model leaves them alone.
Why it matters
Two things change here. The first is speed. OpenAI's stated goal is to take advanced security research to organisations of every size, independent researchers and open-source maintainers, through its Daybreak program. Daybreak Blue helps verified defenders check vulnerabilities and fixes; Daybreak Red, where GPT-6 Cyber sits, allows deeper exploit development and authorised testing. If that works, the software your business runs will be checked more often and patched more often.
The second is the gap between a fix being released and a fix being installed. Attackers read security advisories too. The Essential Eight maturity model from the Australian Signals Directorate says the attackers that Maturity Level One is designed to stop are mostly opportunists: they look for any victim with a common weakness, such as an online service that has not been patched against a public exploit. A long advisory like Firefox's is a to-do list for them as well as for you.
OpenAI is also careful about who gets the model. It requires IP allowlisting for Daybreak Red API access by default, which means requests are accepted from approved networks with limited exceptions, and it monitors credentials for compromise. API customers using Zero Data Retention generally must use Private Safety Processing, which reviews encrypted records kept in the customer's own storage for 30 days. OpenAI also recommends sandboxing, action review and human oversight for high-risk work, the same kind of controls we discussed in our post on CrowdStrike's kill switch for rogue AI agents.
What this means for businesses
You are unlikely to use GPT-6 Cyber yourself. Security teams at Cisco, Calif and CodeRabbit are named as early users, and OpenAI says its mainline models such as GPT-6 Astra already support code review, vulnerability triage and patch development. What lands on your desk is the result: more advisories, more fixes, and less time before someone tries the weaknesses those fixes close.
So the useful question is how fast your updates go in. The Essential Eight's Maturity Level One sets clear targets, explained in our guide to its 48-hour patching rule:
- Web browsers and their extensions, office suites, email clients, PDF software and security products: patched within two weeks of release.
- Online services and the operating systems of internet-facing servers and network devices: within 48 hours when a flaw is critical or a working exploit exists, otherwise within two weeks.
- Operating systems of office computers and internal servers: within one month.
- Browsers and other listed software that the vendor no longer supports: removed.
A short checklist for this month:
- Check that every computer that uses Firefox is on version 159 or later, and that automatic updates are on for every browser, not just the one people use most.
- Ask whoever maintains your website and servers how they hear about advisories and how quickly they apply critical patches. business.gov.au notes that with a self-hosted website you must apply the latest security patches yourself, while a cloud-based platform provider should keep the site safe.
- Run a vulnerability scanner, or ask your IT provider for the latest report, so missing patches show up before an attacker finds them.
- If you commission custom software, ask the developer how they review code for security and how fixes reach you once the system is live.
- Write down who is responsible for patching each system, including the cloud services you rent. OpenAI's earlier misalignment reports showed test models getting past common weak spots on websites.
If nobody owns this work in your business, our managed IT support keeps devices and software patched on a schedule, and our cyber security service helps you apply the Essential Eight in a way that suits a small office. If you are hiring someone in-house to own this work, Talent Tabloid lets you check a candidate's verified work history. You can ask us for a free first consultation and a written quote.
Frequently asked questions
Can a small business sign up for GPT-6 Cyber?
Not straight away. OpenAI says GPT-6 Cyber is first available through the API and Codex to a select group of approved Daybreak Red customers, with access expanding to all Daybreak Red customers over time. New organisations can apply for Daybreak, but most small businesses will meet this work through their security provider rather than use it directly.
Does GPT-6 Cyber make my business safer on its own?
No. It helps security researchers and software makers find and fix flaws faster, which shows up for you as more frequent security updates. Those updates only protect you once they are installed on your devices, servers and website, so your patching routine is what turns the research into protection.
Which Firefox version has the fixes Mozilla released?
Mozilla Foundation Security Advisory 2026-118, announced on 27 October 2026, lists the vulnerabilities fixed in Firefox 159, rated high impact. Check that every computer in your office runs Firefox 159 or later and that automatic updates are switched on.
What does "Critical" mean in OpenAI's rating of GPT-6 Cyber?
It is OpenAI's own classification under its Preparedness Framework, the internal scale it uses to judge how capable and risky a model is in an area. GPT-6 Cyber is rated Critical in cybersecurity, up from High for GPT-5.6 Cyber, which is why OpenAI keeps access behind an application process and extra safeguards.
How fast should a small office install browser updates?
The Essential Eight maturity model from the Australian Signals Directorate says that at Maturity Level One, patches for web browsers and their extensions are applied within two weeks of release. Online services and internet-facing servers have a tighter 48-hour window when a flaw is critical or a working exploit exists.
Key takeaways
- OpenAI launched GPT-6 Cyber on 5 November 2026 and rates it Critical in cybersecurity, up from High for GPT-5.6 Cyber.
- It produced evidence for 81 novel high-severity Firefox vulnerabilities; Mozilla's fixes are in Firefox 159.
- Access is limited to approved Daybreak Red customers, with IP allowlisting on by default.
- For small businesses the change is more patches, sooner: the Essential Eight asks for browser patches within two weeks and critical internet-facing fixes within 48 hours.
- Name an owner for patching every system, from browsers to the website.