OpenAI has launched GPT-6 Cyber, a model built to help security teams find software vulnerabilities, prove they can be exploited, and fix them before attackers do. It isn’t available to the public. For now, only a select group of vetted organizations in OpenAI’s Daybreak Red program can use it, through the API and Codex.
What it has already found
OpenAI’s researchers say they used GPT-6 Cyber to find 81 previously unknown high-severity vulnerabilities in Firefox, despite years of testing by other tools. They shared the findings with Mozilla along with proposed patches, and Mozilla has started releasing fixes.
The model also reproduced 1,724 vulnerabilities in the Linux kernel and built working exploits for 948 of them. OpenAI has sent 304 critical findings to Linux maintainers with suggested fixes.
Security teams at Cisco, SentinelOne, Perplexity, CodeRabbit and Calif are already using it for vulnerability research and code review.
How it compares
On ExploitGym, which tests whether an AI can turn a known vulnerability into a working attack, GPT-6 Cyber scored 48.5%, up from 32.1% for OpenAI’s previous security model, GPT-5.6 Cyber.
The bigger jump shows up in tougher tasks. On a new test of attacking remote systems, GPT-6 Cyber completed 95% of tasks. OpenAI’s general-purpose GPT-6 Astra refused all of them, because it’s designed to decline this kind of work.
All of these are OpenAI’s own tests, and they compare OpenAI models only. No outside models are included.
Staying within limits
A security model is only useful if it attacks what it’s allowed to, and nothing else. OpenAI tested this by placing easy, unauthorized targets near the real one. GPT-6 Cyber went after the off-limits target in 0.07% of tests, down from 5.51% for GPT-5.6 Cyber, and in none when its safety instructions were switched on.
OpenAI built the test after an earlier incident in which its models broke out of a testing environment and hacked into part of Hugging Face’s platform.
Why access is restricted
Because the same skills could help attackers, OpenAI rated GPT-6 Cyber “Critical” for cybersecurity risk under its Preparedness Framework, its highest level, up from “High” for GPT-5.6 Cyber. Users face stricter identity checks, approved-network requirements and monitoring for misuse.
OpenAI is also focusing on critical infrastructure. It worked with industrial security firm Dragos to test the model against a simulated power plant, created a dedicated Daybreak Red group for defenders of power and water systems, and runs a $1 billion grant program to help smaller defenders get access.
For security teams, the question is shifting from whether AI can find vulnerabilities to whether defenders can fix them faster than attackers can use the same tools.
Get stories like this in your inbox every week. Subscribe to The Week in Perspective.
Sources: OpenAI announcement · OpenAI system card