OpenAI unveiled GPT-6 Cyber on Nov. 5, 2026, introducing a cybersecurity research model designed to find and validate exploitable vulnerabilities and help developers fix them. The company said the model improves vulnerability discovery while adhering more closely to authorized testing boundaries, addressing the risk that cybersecurity agents could target third-party services or exceed their assignments.

In its announcement, OpenAI said security teams at Mozilla, Palo Alto Networks, Calif and Cisco were already using GPT-6 Cyber alongside its own researchers.

Firefox and Linux research

OpenAI said GPT-6 Cyber helped identify 81 previously unknown Firefox vulnerabilities that Mozilla classified as critical in severity. The model also indicated that 30 known issues were more serious than previously believed. Researchers shared the findings and proposed fixes with Mozilla, which began rolling out updates.

Brian Grinstead, a distinguished engineer at Firefox, said: “Firefox entered 2026 as one of the world’s most hardened and scrutinized codebases. Even after further hardening with multiple generations of frontier models, GPT‑6 Cyber uncovered previously unknown high-severity vulnerabilities, identified architectural improvements, and helped us ship fixes to millions of users.”

In research involving the Linux kernel, the model reproduced 1,724 vulnerabilities and developed 948 working exploits, OpenAI said. GPT-5.6 Sol exploited 166 vulnerabilities under similar conditions. Researchers shared 304 critical findings and potential fixes with Linux maintainers.

GPT-6 Cyber and other OpenAI models also helped develop and validate exploits for a critical vulnerability in Brotli, a data-compression library used by Node.js and other software. OpenAI said the vulnerability could potentially permit remote code execution in applications that depend on the library.

Benchmark and safety results

GPT-6 Cyber scored 48.5% on ExploitGym, which measures a model’s ability to exploit a specified vulnerability in a remote system, compared with 32.1% for GPT-5.6 Cyber. That represents a relative improvement of about 51%.

OpenAI also tested whether the model would remain within an authorized assignment. In the ExploitGym Honeypot test, GPT-6 Cyber worked in an isolated environment containing additional intentionally vulnerable services. Without special safety instructions, it refrained from exploiting resources outside the assigned task in 99.93% of cases, compared with 94.49% for GPT-5.6 Cyber. With deployment instructions, GPT-6 Cyber followed the constraints in 100% of tests.

The company said it added monitoring mechanisms, suspicious-activity checks and credential protections. Access to the Daybreak Red API requires IP allowlisting by default, while separate restrictions apply to work involving critical infrastructure, including energy facilities, water-treatment systems and industrial control systems.

OpenAI worked with industrial cybersecurity company Dragos to validate the safeguards. In a controlled environment, GPT-6 Cyber identified two zero-day vulnerabilities in engineering software used to manage protective-relay data.

Limited initial access

GPT-6 Cyber initially will be available through the API and Codex to a limited group of approved Daybreak Red customers, according to OpenAI. The company plans to expand access gradually to other participants. It is also developing Daybreak Blue, which will provide tools to vetted system defenders, researchers and security teams.

OpenAI previously launched EVMbench with Paradigm and OtterSec to assess agents’ ability to find and exploit vulnerabilities in Ethereum smart contracts. The announcement also follows demonstrations in which an experimental Anthropic Claude model identified and chained vulnerabilities for complex attacks on Chrome, Safari, Android, Windows and Linux.

JPMorgan Chief Executive Jamie Dimon and Anthropic Chief Executive Dario Amodei have called AI-related cyber threats a serious challenge for the financial sector. Google also reported patching 52 vulnerabilities in Chrome 156, including some discovered using Claude and OpenAI Codex Security models.

Source: Incrypted