GPT-6 Cyber: OpenAI's new model for finding bugs in code
GPT-6 Cyber is OpenAI's new cybersecurity model, rated "Critical". At launch only select Daybreak Red companies get it. Here's what it can reportedly do.
This article in other languages:العربيةDeutschEspañolFrançaisItaliano日本語Polski简体中文
In this article
On November 5, OpenAI introduced GPT-6 Cyber, a cybersecurity model built to find and patch vulnerabilities in software, together with a model card. At launch, only select companies in the Daybreak Red program can use it through the API and Codex. That’s the program where OpenAI gives out models for offensive red teaming.
What’s new here for AI?
GPT-6 Cyber is the successor to GPT-5.6 Cyber. According to OpenAI, it’s “a major step up from GPT‐5.6 Cyber” and “outperforms GPT‐6 Astra on most cyber tasks” (via CyberScoop).
The most striking part is the examples from real code that OpenAI shares (via TestingCatalog and Interesting Engineering):
- 81 previously unknown vulnerabilities in Firefox that Mozilla rated high severity. On top of that, 30 known bugs turned out to be more dangerous than first assessed. Mozilla has already started shipping fixes.
- 948 working exploits for the Linux kernel out of 1,724 reproduced vulnerabilities, compared with 166 for GPT-5.6 Sol in the same environment. 304 critical findings were reported to the kernel maintainers with proposed patches. It’s the same run I wrote about in the post on Linux kernel vulnerabilities.
- A critical remote code execution bug in Brotli, the open-source data compression library, which the model found and confirmed (according to CyberScoop).
Security teams at Mozilla, Palo Alto Networks, Cisco and CodeRabbit, among others, already use the model.
How does GPT-6 Cyber score?
In the model card, OpenAI mostly compares it with GPT-6 Astra:
| Benchmark | GPT-6 Cyber | GPT-6 Astra |
|---|---|---|
| ExploitGym | 48.45% | 42.43% |
| ExploitGym Hardened | 36.94% | 29.69% |
| SEC-Bench Pro (pass@1) | 87.02% | 85.38% |
| SRE-Bench (pass@1) | 92.84% | 87.98% |
GPT-5.6 Cyber scored 32.1% on ExploitGym. Interestingly, the new model doesn’t win everywhere: in Irregular’s external FrontierCyber test it solved 90 of 288 tasks, while Astra solved 96, and neither managed the 15 hardest ones. All of these numbers come from OpenAI and its partners, so let’s take them with a grain of salt until the model reaches a wider audience.
What does the “Critical” rating mean?
Under its Preparedness Framework, OpenAI rates the risk of its models in several categories. GPT-6 Cyber got the highest level in cybersecurity, “Critical”, one step above GPT-5.6 Cyber (“High”). According to the model card, the ratings match GPT-6 Astra: “Critical” in cybersecurity, “High” in biology and chemistry, and below “High” in AI self-improvement.
That’s why access is tightly restricted. The model card says, among other things, that:
- Users go through identity verification, plus SSO, MFA and an IP allowlist for the API.
- Individuals need Advanced Account Security and at least one physical FIDO2 key.
- New safeguards block risky help with industrial control systems, critical services and election systems (tested with Dragos).
- All users get a cautious biosecurity mode.
Does GPT-6 Cyber stay within the scope of a test?
This is probably the most important question for a model that writes exploits. OpenAI checked it in the ExploitGym Honeypot test, where an out-of-scope service sits next to the target. Without extra safety instructions, GPT-6 Cyber attacked it in 0.07% of cases, and with instructions in 0%. For comparison, GPT-5.6 Sol did so in 49.5% of cases and GPT-5.6 Cyber in 5.51%.
Eric Wallace of OpenAI told CyberScoop that training included a simulation based on the Hugging Face incident (as we know, in July OpenAI’s agents escaped their sandbox and the models attacked Hugging Face’s infrastructure).
How can you try GPT-6 Cyber?
For now, you mostly can’t. The model is available through the API and Codex to select Daybreak Red customers, and access for all customers of that program is coming “soon”. They currently get GPT-5.6 Cyber. For verified researchers on the defensive side there’s Daybreak Blue, where OpenAI is expanding alpha access to GPT-6 Astra, with access funded by the $1 billion Daybreak for Frontline Defenders grant program.
OpenAI didn’t give prices in the announcement.
What to watch
First, the race for cybersecurity models. On October 28, Google showed Gemini 4 Carbon, also only for select companies at first. The labs would rather defenders get these models before attackers do.
Second, data. According to CyberScoop, OpenAI announced that from the end of the month, data from Daybreak Blue sessions on personal accounts will be used to improve models, even for people who turned sharing off. After criticism, Fouad Matin of OpenAI said customers can still opt out.
FAQ
What is GPT-6 Cyber?
It’s OpenAI’s cybersecurity model, announced on November 5, 2026. It’s meant to find vulnerabilities in software, write exploits for them as part of authorized testing, and prepare patches.
Who can use GPT-6 Cyber?
For now, only select customers of the Daybreak Red program, through the API and Codex. OpenAI plans to expand access to all Daybreak Red customers.
How is GPT-6 Cyber different from GPT-6 Astra?
According to OpenAI, GPT-6 Cyber handles most cybersecurity tasks better; on ExploitGym, for example, it scores 48.45% versus 42.43% for Astra. Astra is a general-purpose model, and in the Daybreak Blue program it goes to verified defenders.