Recommended Articles
Tech & AI

OpenAI Releases GPT-6 Cyber, Rated "Critical" for Cybersecurity Capability

By Angeled · November 6, 2026 · Tech & AI
OpenAI Releases GPT-6 Cyber, Rated "Critical" for Cybersecurity Capability
OpenAI Releases GPT-6 Cyber, Rated "Critical" for Cybersecurity Capability

OpenAI released GPT-6 Cyber on November 5, 2026, a new model aimed at cybersecurity work, to a small group of approved customers under its Daybreak Red program. The company rated the model "Critical" for cybersecurity capability, a step up from the "High" rating it gave GPT-5.6 Cyber, the previous generation of its security-focused model.

The release is notable less for what the model can do in a demo than for how OpenAI is choosing to distribute it. Rather than a general launch, GPT-6 Cyber is going first to a vetted group of customers, reflecting how seriously the company treats the dual-use nature of a model built to find and exploit software vulnerabilities.

What GPT-6 Cyber Is

GPT-6 Cyber is a specialised variant of OpenAI's GPT-6 family, tuned for cybersecurity tasks such as vulnerability discovery, exploit development, and security analysis. The distinction matters: general-purpose frontier models are trained to be broadly capable across writing, coding, reasoning, and conversation, while a cyber-focused variant is optimised for the narrow but demanding work of understanding software systems deeply enough to find their weaknesses.

Cybersecurity capability in a language model is not a single skill. It spans reading and reasoning about source code, understanding operating system internals, chaining together the small flaws that add up to a working exploit, and writing the proof-of-concept code that demonstrates a vulnerability is real. A model that can do all of this reliably is a powerful tool for defenders — and, in the wrong hands, a powerful tool for attackers. That tension is the central story of this release.

The ExploitGym Benchmark

OpenAI reported that GPT-6 Cyber scored 48.5% on its ExploitGym benchmark, compared with 42.4% for GPT-6 Astra. ExploitGym is a benchmark designed to measure a model's ability to find and exploit vulnerabilities in realistic software environments, rather than simply answering security questions or spotting textbook flaws in isolated code snippets.

Benchmark scores in this range are worth reading carefully. A score of 48.5% means the model succeeds on just under half of the benchmark's tasks — a meaningful improvement over its predecessor's 42.4%, but also a reminder that even the most capable models still fail more often than they succeed at end-to-end exploit development. For founders and investors in the security space, the direction of travel matters more than the absolute number: each generation is closing the gap, and the gap is closing fast.

81 Previously Unknown Firefox Vulnerabilities

Alongside the release, OpenAI reported that the model had found 81 previously unknown high-severity vulnerabilities in Firefox. The finding is a concrete demonstration of the model's capability rather than a theoretical claim. Firefox is one of the most heavily scrutinised codebases in the world, maintained by Mozilla and examined constantly by security researchers. Finding 81 previously unknown high-severity flaws in it is a significant result.

For the browser security ecosystem, the implication is double-edged. On one hand, vulnerabilities found by a model and disclosed responsibly make software safer. On the other, the same capability, in other hands, could be pointed at software that has not yet been patched. The balance between those two outcomes is precisely why OpenAI is limiting initial access.

What "Critical" Means

OpenAI's capability ratings run on a scale that places "Critical" above "High." The upgrade from "High" for GPT-5.6 Cyber to "Critical" for GPT-6 Cyber signals that OpenAI's own internal evaluations found a material jump in the model's ability to perform cyber operations. These ratings are part of the company's preparedness framework, under which higher capability levels trigger stricter deployment controls.

The Daybreak Red program is the mechanism for those controls here. By releasing first to a small group of approved customers, OpenAI can observe how the model is used in practice, gather feedback from trusted security teams, and refine its safeguards before any wider release. It is a pattern the industry has increasingly adopted for capabilities with clear misuse potential: limited, monitored deployment first, general availability later — if at all.

Why It Matters for Founders and Investors

For founders building in cybersecurity, GPT-6 Cyber changes the competitive landscape in both directions. Defensive security startups — vulnerability management, penetration testing, code review, and threat detection — gain a far more capable engine to build on, but they also face a world in which the underlying capability is increasingly commoditised by the model providers themselves. The durable value is likely to sit in the workflow, the integration with a customer's systems, and the trust that comes from responsible handling of findings, rather than in the raw ability to find a bug.

Read Next