OpenAI releases GPT-6 Cyber with a Critical cybersecurity rating

OpenAI released GPT-6 Cyber on Thursday. The model is built to find security flaws and to work out how they could be exploited. Onl...

OpenAI released GPT-6 Cyber on Thursday. The model is built to find security flaws and to work out how they could be exploited. Only a small group of vetted customers can use it for now. OpenAI rated the model "Critical" for cybersecurity under its safety framework, the highest label it has given a cyber-focused release. The previous model, GPT-5.6 Cyber, was rated "High."

The company says the aim is to help security researchers find vulnerabilities, test how they could be exploited and fix them before attackers get there. That goal is not new. Defenders have always tried to find holes before criminals do. What has changed is the tool. A language model can now read code, reason about it and produce working exploits at a pace that a small team of human researchers cannot match.

This article covers what GPT-6 Cyber is, what OpenAI says it has already done, who can use it, and why the "Critical" rating deserves attention.

What is GPT-6 Cyber?

GPT-6 Cyber is a specialized model for security work. OpenAI describes it as a system that helps researchers discover vulnerabilities, test how those vulnerabilities could be exploited and patch them. Finding a bug is only part of the job in security. A report that says "this code might be unsafe" is far less useful than one that shows the flaw can be turned into a working attack. Exploit development is how researchers prove a bug matters, and it is how vendors decide what to fix first.

Giving a model this ability is a sharp choice. The same skill that helps a defender demonstrate a risk could help an attacker create one. That tension explains the limited release and the heavy attention OpenAI has paid to access controls.

Why the "Critical" cybersecurity rating matters

OpenAI uses a safety framework to classify how capable its models are in sensitive areas, and cybersecurity is one of them. GPT-5.6 Cyber was rated "High." GPT-6 Cyber moves up to "Critical."

OpenAI has not framed this as a routine bump. A higher classification means the company sees the model as more capable of causing harm if it were misused, and so it calls for tighter handling. That explains why the launch is restricted to select customers rather than offered broadly. The rating reads as a warning label and also as an admission of how much the model can do.

For readers who follow AI policy, the classification is also useful as a data point. It shows a major developer assessing its own model and concluding that the risk has risen from one generation to the next. Whether the safeguards match that risk is a question outside observers will keep asking.

What OpenAI says the model has already found

OpenAI shared several results from its own testing. They give a rough sense of scale.

Firefox findings

OpenAI said its researchers used the model to find 81 previously unknown vulnerabilities in Firefox. Mozilla classified all 81 as high severity, and it has begun releasing fixes. Browsers are among the most heavily tested pieces of software in everyday use. They process untrusted content from the open web constantly, so they attract a great deal of scrutiny from both defenders and attackers. Turning up 81 new high-severity issues in a target like that suggests the model can find things experienced human reviewers and existing tools missed.

The fact that Mozilla confirmed the severity and started shipping patches matters. It means the findings were not simply claims from OpenAI. A vendor looked at them, agreed they were serious and acted.

Linux kernel findings

The Linux kernel results are larger. OpenAI said the model helped develop working exploits for 948 Linux kernel vulnerabilities. Of those, 588 were previously unknown, which leaves 360 that had already been identified.

The company also disclosed 304 critical Linux findings to maintainers, along with proposed patches. The kernel sits underneath servers, phones, embedded devices and cloud infrastructure, so weaknesses there can ripple outward to a huge number of systems. Sending patches alongside the disclosures is a practical step. Maintainers get a problem and a possible fix together, which can shorten the time between discovery and repair.

These numbers come from OpenAI. Independent confirmation will take time, and kernel maintainers will decide for themselves how many of the proposed patches to accept. Still, the figures show the direction of travel: AI systems are getting good at the tedious, detailed work of auditing very large codebases.

Who can use GPT-6 Cyber: the Daybreak Red tier

GPT-6 Cyber is available to select customers of Daybreak Red, which OpenAI describes as its access tier for advanced, authorized security work. Penetration testing and exploit development are the examples the company gives. In both, a person is hired or approved to attack a system on purpose so its owner can learn where it is weak.

The "authorized" part is the point of the tier. Customers are vetted before they get in, and the work they do is supposed to be sanctioned. OpenAI says it will expand GPT-6 Cyber to all approved Daybreak Red customers over the coming weeks. For now, the rollout is staged, which gives the company room to watch how the model is used before opening it wider.

GPT-5.6 Cyber opens up to all Daybreak Red customers

OpenAI is also making GPT-5.6 Cyber, the earlier model, available to all Daybreak Red customers starting Thursday. This sets up two levels of access. Every approved customer can use the older model right away, while the newer and more capable one reaches a smaller group first and spreads outward later.

For security teams, the older model may be enough for many tasks. Teams doing routine assessments might not need the strongest option, and the staged approach lets OpenAI keep its most powerful tool in fewer hands during the early period.

Growing concern over AI misuse

OpenAI said it is seeing more attempts to misuse AI, including by malicious cyber actors. It also pointed to groups trying to distill its models.

Distillation means training a smaller or separate model on the outputs of a stronger one, so the copy picks up some of the original's abilities. For a model with offensive security skills, that is a real worry. If someone can extract those skills through repeated queries and train their own system, the controls on the original start to matter less. OpenAI's mention of it suggests the company views access restrictions as only one layer of defense, and that it is watching for extraction attempts as well as direct misuse.

The announcement presents the model as a response to this trend. Attackers are already experimenting with AI, the argument goes, so defenders need strong tools of their own. That logic is common in security. It also carries a risk, since building a more capable cyber model adds to the pool of dangerous capability, even when the intent is protective.

What this means for security teams

For organizations that run penetration tests or maintain large codebases, a few practical effects stand out.

Faster discovery: If a model can surface dozens of high-severity browser bugs and hundreds of kernel issues, the backlog of unknown flaws in common software may be bigger than anyone assumed. More discovery means more patching work for vendors and maintainers.

Pressure on patch timelines: When working exploits can be built quickly, the gap between a bug being found and it being usable shrinks. Teams that are slow to apply updates will feel that more.

A new baseline for testing: Penetration testers who have access to these models may be able to cover more ground in the same engagement. Customers who hire testers may start to expect that level of coverage.

Greater need for disclosure discipline: Coordinating with maintainers, as OpenAI says it did with its 304 critical Linux findings, becomes a bigger job when the volume of findings rises. Open-source projects with few volunteers could struggle to keep up.

Open questions and risks

Several things remain unclear from the announcement. It is not yet known how strict the vetting process is, or how OpenAI will detect a customer who uses the model for something other than authorized work. The company's own words show it expects misuse attempts, so enforcement will be tested.

Independent verification is another gap. The Firefox figures carry weight because Mozilla classified the bugs and began fixing them. The Linux numbers, including the 588 previously unknown vulnerabilities and the 304 critical disclosures, rest for now on OpenAI's account. Maintainers' responses over the coming weeks will show how many of those findings hold up.

There is also the broader debate about whether publishing results like these encourages others to build similar systems. Some researchers argue that openness about capability helps defenders prepare. Others worry it speeds up an arms race. The "Critical" label puts OpenAI on record as taking the risk seriously, but it does not settle that argument.

GPT-6 Cyber marks a step in how AI is used for security. It finds flaws, helps build the exploits that prove they are real, and supports fixing them. OpenAI's reported results include 81 high-severity Firefox vulnerabilities, working exploits for 948 Linux kernel vulnerabilities and 304 critical Linux disclosures with proposed patches. Those are large numbers, and Mozilla's response gives at least part of them outside backing.

The "Critical" rating shows OpenAI knows the same power could help attackers. Access through Daybreak Red, a staged rollout and the continued availability of GPT-5.6 Cyber are its answer. Over the next few weeks, as more approved customers get the model, the real test begins: whether the benefits for defenders outweigh the risks that come with such a capable tool.

Frequently Asked Questions (FAQ)

What is GPT-6 Cyber?

GPT-6 Cyber is an OpenAI model built to find security vulnerabilities, test how they could be exploited and help fix them. It is designed for authorized security research.

Why did OpenAI give GPT-6 Cyber a "Critical" rating?

Under its safety framework, OpenAI raised the model's cybersecurity classification from "High," the level of GPT-5.6 Cyber, to "Critical." The higher label reflects greater capability and a greater potential for harm if the model were misused.

Who can use GPT-6 Cyber right now?

Only select customers of Daybreak Red, OpenAI's access tier for advanced, authorized security work. OpenAI plans to extend access to all approved Daybreak Red customers over the coming weeks.

What is Daybreak Red?

Daybreak Red is OpenAI's access tier for authorized security work such as penetration testing and exploit development. Customers are vetted before they can use the models offered through it.

What did GPT-6 Cyber find in Firefox?

OpenAI said its researchers used the model to find 81 previously unknown Firefox vulnerabilities. Mozilla classified them as high severity and has started releasing fixes.

What did the model do with Linux kernel vulnerabilities?

OpenAI said the model helped develop working exploits for 948 Linux kernel vulnerabilities, including 588 that were previously unknown. The company also disclosed 304 critical Linux findings to maintainers, with proposed patches.

Is GPT-5.6 Cyber still available?

Yes. OpenAI is making GPT-5.6 Cyber available to all Daybreak Red customers starting Thursday.

What is model distillation, and why does OpenAI mention it?

Distillation is the practice of training another model on a stronger model's outputs to copy some of its abilities. OpenAI said it is seeing groups try to distill its models, which could let others reproduce capabilities without going through its access controls.

Can attackers use GPT-6 Cyber?

OpenAI limits access to vetted customers doing authorized work, and it says it is seeing more attempts to misuse AI. The restrictions are meant to keep the model away from malicious actors, though how well they work will only become clear with time.

References

Reuters - https://www.reuters.com/technology/openai-launches-gpt-6-cyber-with-critical-cybersecurity-rating-2026-11-05/

Open ai - https://openai.com/index/path-to-astra/

Loaded All Posts Not found any posts VIEW ALL Readmore Reply Cancel reply Delete By Home PAGES POSTS View All RECOMMENDED FOR YOU LABEL ARCHIVE SEARCH ALL POSTS Not found any post match with your request Back Home Sunday Monday Tuesday Wednesday Thursday Friday Saturday Sun Mon Tue Wed Thu Fri Sat January February March April May June July August September October November December Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec just now 1 minute ago $$1$$ minutes ago 1 hour ago $$1$$ hours ago Yesterday $$1$$ days ago $$1$$ weeks ago more than 5 weeks ago Followers Follow THIS PREMIUM CONTENT IS LOCKED STEP 1: Share to a social network STEP 2: Click the link on your social network Copy All Code Select All Code All codes were copied to your clipboard Can not copy the codes / texts, please press [CTRL]+[C] (or CMD+C with Mac) to copy Table of Content