OpenAI introduced GPT-6 Cyber on November 5, a specialized model designed to find, validate, and help fix software vulnerabilities. The company describes it as its most advanced cybersecurity system to date, but it is not opening the model to the general public. Initial access will be available through the API and Codex to a selected group of approved Daybreak Red customers, OpenAI’s program for trusted offensive and defensive security teams.[1]

The launch combines a technical jump with a test of control. According to OpenAI, GPT-6 Cyber scored 48.5% on ExploitGym, an evaluation that measures whether a model can exploit a specified vulnerability, compared with 32.1% for GPT-5.6 Cyber. That represents a 51% relative improvement. The company also says the new model avoided exploiting services outside the authorized scope in 99.93% of a separate evaluation without its deployment prompt, and in 100% of cases when that prompt was active.[1]

Those figures come from OpenAI’s own evaluations and do not, by themselves, amount to an independent audit. The Verge confirmed the central elements of the announcement and emphasized the promise that the system will keep research within boundaries set by the user, but the technical results and disclosed findings still rely primarily on data published by the company.[2]

OpenAI connects the model to work performed on real projects. It says its researchers used GPT-6 Cyber to develop 948 Linux privilege-escalation exploits and identify 81 previously unknown high-severity vulnerabilities in Firefox. The company also says the system has helped uncover and validate thousands of exploits in security-critical codebases. The Verge highlights the Firefox case and reports that the model also played a role in hardening OpenAI’s own infrastructure for training AI systems.[1][2]

The wording requires care. Finding or building an exploit in an authorized environment does not mean the flaw was exploited outside the lab, and it does not establish that maintainers accepted every result as a separate vulnerability. OpenAI presents the work as coordinated research intended to fix software. Restricted availability also indicates that the company believes these offensive capabilities require additional controls even when their stated purpose is defensive.

OpenAI says those controls include dedicated training to respect authorized scope, access restrictions, and monitoring intended to detect and disrupt misuse. The Verge also reports new checks designed to block high-risk activity, including attacks against critical infrastructure such as power plants and water-treatment facilities.[2] OpenAI says it will progressively expand GPT-6 Cyber access to all Daybreak Red customers. In the meantime, GPT-5.6 Cyber will be available to every participant in that program.[1]

The launch follows OpenAI’s September designation of Astra as its first model to reach the “Critical” cybersecurity capability threshold under the company’s Preparedness Framework. GPT-6 Cyber turns that broader discussion about capability and safeguards into a specialized product. It is not merely a general assistant with an added security feature; it is a model oriented toward security review, vulnerability validation, and red-team work within defined boundaries.

The most important element is not a single benchmark number but the operational shift it represents. AI agents are no longer limited to suggesting code; they can chain together inspection, testing, and vulnerability validation. The unresolved question is whether scope controls, customer screening, and monitoring will maintain the same performance as access expands. For now, the reported metrics are promising, but the strongest public evidence still comes from OpenAI. Independent validation and reports from software maintainers will be needed to measure the model’s real-world impact.

Sources

[1] Read More — GPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents — OpenAI
[2] Read More — OpenAI says its new cybersecurity model is better at not hacking other companies — The Verge