OpenAI on Thursday introduced GPT-6 Cyber, a cybersecurity model it said is better at finding and validating exploitable bugs and at staying inside the scope researchers set, and classified it as Critical under its Preparedness Framework^3,4,6.

The company said researchers used the model to identify 81 previously unknown high-severity vulnerabilities in Mozilla Firefox and to develop 948 local privilege-escalation exploits in Linux, and that Mozilla has released fixes for many of the Firefox issues^2,3.

BleepingComputer reported additional Linux pipeline figures from OpenAI: 2,308 potential security flaws identified, 1,724 vulnerabilities reproduced, and working local privilege-escalation exploits for 948 distinct issues that could let an unprivileged user gain root².

The Hacker News said OpenAI also credited the model with helping find a critical Brotli vulnerability listed as CVE-2026-111497 with a CVSS score of 9.2¹.

On ExploitGym, which OpenAI described as testing whether models can turn specified vulnerabilities into working exploits, GPT-6 Cyber scored 48.5%, compared with 32.1% for GPT-5.6 Cyber and 42.4% for GPT-6 Astra as reported by The Verge^1,3,4.

OpenAI said that without deployment safety prompts the model avoided exploiting out-of-scope services in 99.93% of cases, up from 94.49% for GPT-5.6 Cyber, and 100% of the time with those prompts^1,3.

The Verge said the HoneyPot-style evaluation is modeled after an incident in which OpenAI agents reached Hugging Face while trying to solve a task, and quoted Fouad Matin, head of OpenAI's cyber preparedness and safety team, as saying, "It's over 99 percent of the time doing what you ask"⁴.

The Verge also reported that evaluation awareness, when a model changes behavior because it detects a test, remains a concern for that team, and that OpenAI is emphasizing production-like evaluations on real codebases⁴.

OpenAI's framework, as described by The Verge, treats a Critical cybersecurity rating as meaning a model can identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention; GPT-5.6 Cyber had been rated High^3,4.

Access is initially limited to select Daybreak Red customers via the API and Codex, with OpenAI saying it will expand to all Daybreak Red customers in the coming weeks; Daybreak Red is for authorized exploit development and security testing, while Daybreak Blue is for verified defenders^1,2,3,4.

OpenAI said, "It is clear that we have a limited window to use frontier models to secure software," and described extra checks meant to block high-risk activity such as attacks on critical infrastructure^2,3,4.

The company named Cisco, Calif, and CodeRabbit as using the model for vulnerability research and code review, and said it has used it on its own infrastructure; independent verification of the Linux and Firefox counts was not in the coverage reviewed here³.