OpenAI opens GPT-6 Cyber to vetted security teams, citing 81 Firefox flaws

OpenAI is giving approved security teams access to GPT-6 Cyber, a model designed to find software vulnerabilities and demonstrate whether they can actually be exploited, according to a November 10 report from EdTech Innovation Hub. Access runs through Daybreak Red, OpenAI's program for more advanced exploit development and authorized security testing.

The most striking results came from Firefox. OpenAI says researchers used the model to uncover 81 previously unknown vulnerabilities that Mozilla classified as high severity, and Mozilla has released fixes for many of them after earlier rounds of testing with other advanced AI models.

Brian Grinstead, a Distinguished Engineer at Firefox, said in the report: “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.” The emphasis is on finding new flaws after repeated hardening, not merely on counting bugs.

The Linux kernel numbers are larger. OpenAI says its researchers used GPT-6 Cyber to develop working exploits for 948 Linux kernel vulnerabilities, and that it disclosed 304 critical findings to maintainers with proposed fixes, following Linux community guidance, while work on the rest continues. Cisco, Calif, CodeRabbit and Dragos are among the security teams named as users of the model for vulnerability research and code review.

On capability benchmarks, GPT-6 Cyber scored 48.5% on OpenAI's ExploitGym evaluation, which tests whether a model can exploit a specified vulnerability, compared with 32.1% for GPT-5.6 Cyber. OpenAI also raised the model's cybersecurity capability rating to Critical under its Preparedness Framework, up from High for its predecessor.

That Critical rating sits naturally alongside restricted access: the more reliably a model can turn a flaw into working attack code, the less sense it makes to release it without conditions. Initial access is limited to approved Daybreak Red customers, and OpenAI says the work complied with authorized testing boundaries.

The broader signal is that frontier models are moving from helping find bugs to proving exploitability. That is an efficiency gain for defenders and a capability multiplier for attackers, which is why the release mechanism — who gets access and under what limits — has become part of the product.

Three things are worth watching: how Mozilla and Linux maintainers handle the remaining flaws and fix cadence, whether OpenAI widens Daybreak Red eligibility, and whether the Critical classification brings stricter deployment rules or outside audit requirements.