OpenAI Introduces GPT-6 Cyber: What It Can Do

OpenAI has introduced GPT-6 Cyber — a specialized artificial intelligence model for finding, verifying, and fixing vulnerabilities in software. 

This was reported by OpenAI.

Introducing GPT‑6 Cyber: a major advance in security research that pairs stronger capabilities with better model alignment. The model powered the Linux kernel research we announced yesterday, and security teams at companies like Mozilla, Palo Alto Networks, Calif, and Cisco—alongside our own at OpenAI—are already putting it to work. GPT-6 Cyber is available to select Daybreak Red customers and we are expanding access over time. https://t.co/Nzab3Z1Td6

— OpenAI (@OpenAI) November 5, 2026

The company says the new model performs cybersecurity tasks better than its predecessor, GPT-5.6 Cyber, and also adheres more precisely to the constraints set by researchers.

What GPT-6 Cyber can do

GPT-6 Cyber is a specialized version of GPT-6 Astra trained to analyze source code, find vulnerabilities, verify exploitability, and help developers fix issues.

In particular, the model can examine large software projects, identify potential weak points, and create exploits — programs or sequences of actions that demonstrate how an attacker could abuse a vulnerability. This allows security specialists to assess how dangerous a discovered issue is and determine what fixes are needed to protect a system.

According to OpenAI’s testing on the ExploitGym benchmark, which evaluates models’ ability to exploit given vulnerabilities, GPT-6 Cyber completed 48.5% of tasks versus 32.1% for GPT-5.6 Cyber. That’s roughly a 51% better result on a relative basis.

In another test the model successfully completed 95% of tasks that involved exploiting known vulnerabilities in software such as GitLab, PostgreSQL, and Jenkins.

Which vulnerabilities the model found in Firefox and Linux

During the Firefox research the model helped discover 81 previously unknown vulnerabilities that Mozilla classified as high-severity issues. Researchers also determined that another 30 known vulnerabilities were more dangerous than previously believed.

The company also audited the Linux operating system kernel. In that research the model was able to reproduce 1,724 vulnerabilities and develop working exploits for 948 of them. These exploits demonstrated how a limited-privilege user could gain elevated privileges on the system.

Additionally, the company used GPT-6 Cyber to test its own infrastructure. In controlled tests, AI agents were able to find vulnerability chains that allowed them to escape isolated virtual environments and execute code on host servers.

Will OpenAI restrict use of GPT-6 Cyber?

The company paid special attention to ensuring the model performs only permitted tasks and does not attempt to access third-party systems. In one test GPT-6 Cyber was required to find a vulnerability in a specified system without attacking other available services.

Without additional protective instructions the model respected these limits in 99.93% of cases. With supplementary instructions the success rate was 100% within the testing scope.

For comparison, GPT-5.6 Cyber without such instructions stayed within task boundaries in 94.49% of cases.

At the same time OpenAI classified GPT-6 Cyber as a critical-level cybersecurity capability under its own Preparedness Framework risk assessment system. This is due to its ability to find and exploit complex vulnerabilities.

To reduce the risk of abuse, the company implemented additional monitoring of model usage, checks on permitted actions, and restrictions on working with critical infrastructure.

Currently GPT-6 Cyber is available only to a limited set of vetted users in the Daybreak Red program via the API and Codex. OpenAI plans to expand access to other program participants over time.

Recall that in October 2026 ChatGPT introduced the Locked Chats feature, which allows moving individual conversations from the regular history and search into a protected section. New messages in protected chats will not update the memory the AI uses in regular conversations.

Read also: OpenAI will merge ChatGPT and Codex limits: what will change for users