On 5 November 2026 OpenAI introduced GPT-6 Cyber, a version of GPT-6 trained specifically for security research: finding vulnerabilities, proving they can be exploited, and helping write fixes. It is not available in ChatGPT. Access is limited to vetted security teams through OpenAI's Daybreak programme. It is the second model of this kind in two weeks, after Google's Gemini 4 Carbon, and it confirms a trend: the most capable AI models are now being released to defenders first.
The benchmark results
Source: OpenAI, 5 November 2026. GPT-6 Astra with standard safeguards refuses these tasks, so it scores 0%. Daybreak Blue is OpenAI's access tier for verified defenders.
On a new test of web penetration testing without access to source code, GPT-6 Cyber found 40.3% of vulnerabilities against 32.0% for GPT-6 Astra. On another new benchmark, RemoteExploitBench, which uses known flaws in real applications such as GitLab, PostgreSQL and Jenkins, GPT-6 Cyber completed 95% of tasks while GPT-6 Astra refused them all.
What it has already found
GPT-6 Cyber in the real world
948 working exploits
It reproduced 1,724 kernel vulnerabilities and wrote working privilege-escalation exploits for 948, over five times the 166 achieved by GPT-5.6 Sol with the same setup. OpenAI disclosed 304 critical findings to maintainers.
81 new high-severity bugs
Plus 30 known issues shown to be more severe than reported. Mozilla has shipped fixes for many of them.
Remote code execution shown
A compression library used by Node.js and other software.
Security teams in productionNow
Named users include Mozilla, Cisco, CrowdStrike, Palo Alto Networks, SentinelOne, Perplexity and CodeRabbit.
Staying inside the lines
The risk with a security agent is not only what it can do but whether it does only what it was asked. OpenAI built a test called ExploitGym Honeypot: the model gets a hard task plus some easy-to-attack services that are explicitly out of scope. GPT-6 Cyber left the out-of-scope services alone in 99.93% of runs without its deployment prompt (GPT-5.6 Cyber: 94.49%) and 100% with it. OpenAI also requires IP allowlisting for Daybreak Red API access, monitors for stolen credentials and adds checks against attacks on critical infrastructure such as power and water systems.
GPT-6 Cyber vs Gemini 4 Carbon
GPT-6 Cyber
OpenAI, 5 November- Built for
- finding, exploiting and fixing vulnerabilities
- Access
- selected Daybreak Red customers through the API and Codex, expanding to all Daybreak Red customers
- Headline
- 48.5% ExploitGym, 948 Linux exploits
- Specialised model with dedicated alignment training for staying in scope
A tool for security teams, not a general assistant.
Gemini 4 Carbon
Google, 28 October- Built for
- a general frontier model, released to cyber defenders first
- Access
- cyber defenders, then the Fairwind partner programme
- Headline
- top of the Artificial Analysis index at 59.4; 71% on CWE-bench
- Also the strongest general model on public tests
Expected to become a general-purpose model later.
What this means for data teams
You will probably never use GPT-6 Cyber yourself. But the systems data teams depend on, such as PostgreSQL, Jenkins, GitLab, notebook servers and open-source libraries, are exactly what these models are getting good at attacking. OpenAI talks about a "limited window" in which defenders can use these models before attackers have similar ones. Practical steps:
- 1
Patch faster
Shorten the gap between a security release and your upgrade, especially for databases, CI servers and anything exposed to the internet.
- 2
Know what you run
Keep a list of the versions of PostgreSQL, Spark, Airflow, Jenkins and Python libraries in production, so you can act on advisories quickly.
- 3
Remove public exposure
Put database ports, notebook servers and admin consoles behind a VPN or private network.
- 4
Give agents narrow permissions
If you use AI agents on your data platform, follow our checklist for securing AI agents.
- 5
Use AI on your own code
Mainline models already help with code review and finding bugs; ask them to review your pipelines and SQL for injection and secrets.
Check your understanding
1/3Where can you use GPT-6 Cyber?
Access starts with selected Daybreak Red customers through the API and Codex.
2/3What did GPT-6 Cyber score on ExploitGym?
That is a 51% relative improvement over GPT-5.6 Cyber.
3/3What is the main lesson for teams running databases and CI servers?
Models like this shrink the time between discovery and exploitation.
You got 0 out of 3.
Related: Gemini 4 Carbon, explained, the best AI models in November 2026 and when AI agents go off script.
Sources: OpenAI, "GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents" (5 November 2026) and its system card; Fortune and TechCrunch coverage (5 November 2026); Google DeepMind and Artificial Analysis on Gemini 4 Carbon (28 October to 3 November 2026).
Alessandro Ecclesie AgazziWritten by Alessandro Ecclesie Agazzi, analytics engineer in London. Updated 9 November 2026. About me