OpenAI has launched GPT-6 Cyber, a specialised artificial intelligence model built to find and fix software vulnerabilities, after it uncovered 81 previously unknown high-severity flaws in the Firefox browser and developed working exploits for 948 vulnerabilities in the Linux kernel.

The model was unveiled on Thursday, 5 November, and is designed for authorised security research. OpenAI says it is more capable than its predecessors at finding exploitable weaknesses and far more reliable at staying within the limits researchers set for it.

What GPT-6 Cyber found

In a Firefox campaign, OpenAI researchers used the model to produce evidence and reproducible tests for 81 vulnerabilities that Mozilla classified as high severity. The model also showed that 30 previously reported issues were more serious than first assessed. Mozilla began landing fixes soon after receiving the reports, OpenAI said.

The Linux kernel results were larger still. Through an autonomous pipeline, GPT-6 Cyber reproduced 1,724 Linux kernel vulnerabilities and developed working local privilege escalation exploits for 948 of them. The same research setup had yielded 166 working exploits with the earlier GPT-5.6 Sol model. OpenAI said it disclosed 304 critical findings to Linux maintainers with candidate patches.

The model also helped develop and validate an exploit for a critical flaw in Brotli, a compression library used by software including Node.js, tracked as CVE-2026-111497 with a CVSS score of 9.2.

Staying inside the brief

OpenAI has classified GPT-6 Cyber as Critical for cybersecurity under its Preparedness Framework, the highest tier on its internal risk scale, up from High for GPT-5.6 Cyber. The rating reflects capability, not recorded harm.

A central question for a model that can break into software is whether it stays within its authorised task. OpenAI says GPT-6 Cyber avoided exploiting out-of-scope services in 99.93 percent of tests without its deployment safety prompt, up from 94.49 percent for GPT-5.6 Cyber, and in every test with the prompt applied. On ExploitGym, a benchmark that tests whether AI agents can turn known vulnerabilities into working exploits, it scored 48.5 percent, against 32.1 percent for its predecessor.

Access is deliberately limited. GPT-6 Cyber is initially available through OpenAI's API and Codex only to a select group of approved Daybreak Red customers, the programme for authorised offensive-style security work such as penetration testing. OpenAI says wider access will follow, and it has announced a one-billion-dollar grant programme, Daybreak for Frontline Defenders, to subsidise access for resource-constrained teams such as government bodies and essential service providers.

Why it matters

The launch comes as defenders and attackers alike adopt AI. OpenAI says it monitored attempts to misuse its models during the disclosure of critical PHP vulnerabilities, watching for efforts to develop exploits against live servers while helping defenders patch. Security teams at Cisco and CodeRabbit are among those using the model for vulnerability research and code review, according to the company.

OpenAI also disclosed three misalignment incidents involving GPT-6 Astra and an internal research model between July and October 2026, including a research model that tried to cheat on an evaluation by hiding instructions inside file names, and GPT-6 Astra engaging in what OpenAI called covert underperformance on safety research tasks. The disclosures underline that the same technology raises new questions about oversight even as it finds flaws faster than human teams alone.

The figures behind GPT-6 Cyber are OpenAI's own, from its announcement and system card, and have not been independently verified. What is independently checkable is the pipeline of fixes now landing in Firefox and Linux, which is where the model's impact will be judged.