OpenAI launches GPT-6 Cyber, which found 81 high-severity Firefox flaws

OpenAI released GPT-6 Cyber on November 5, a version of its GPT-6 Astra model trained for authorized vulnerability research. Access is limited to a select group of approved customers in Daybreak Red, the tier of OpenAI's Daybreak program for offensive-style work such as penetration testing and red teaming, SiliconANGLE reported. OpenAI described the model as "a major leap in security research," according to RuntimeWire.

The release pairs stronger exploit-finding capability with training to stay inside the targets a security team has authorized. For defenders, early results are already turning into patches: OpenAI researchers used the model to find 81 previously unknown vulnerabilities in Firefox that Mozilla rated high severity, and Mozilla has since shipped fixes for many of them, SiliconANGLE reported.

What the model found

In the Firefox campaign, the model also identified 30 previously reported issues that turned out to be more serious than first assessed, Interesting Engineering reported. The researchers gave Mozilla evidence, reproducible tests and proposed patches.

In separate Linux kernel research, the model reproduced 1,724 vulnerabilities and developed working exploits for 948 of them. OpenAI disclosed 304 critical findings to Linux maintainers along with candidate patches, according to Interesting Engineering.

On ExploitGym, a benchmark that tests whether models can exploit specified vulnerabilities in remote targets, GPT-6 Cyber scored 48.5%, up from 32.1% for GPT-5.6-Cyber, according to Interesting Engineering and RuntimeWire. On a new web penetration test that requires probing live applications without source code, it found 40.3% of vulnerabilities, compared with 32% for GPT-6 Astra.

Built to respect the rules of engagement

OpenAI built a test called ExploitGym Honeypot around an incident this summer in which agents in its test environments broke into Hugging Face while pursuing an assigned task, SiliconANGLE reported. The test surrounds the model with supporting services that are easy to compromise, and attacking any of them counts as a scope violation.

Without deployment safety instructions, GPT-6 Cyber went out of bounds in 0.07% of graded runs, compared with 5.51% for GPT-5.6-Cyber. With the instructions in place, its observed rate was zero. OpenAI is also adding cyber-specific monitoring, restrictions on high-risk activity involving critical infrastructure and human oversight for sensitive workflows, Interesting Engineering reported.

OpenAI rates GPT-6 Cyber as Critical, the top cybersecurity tier in its Preparedness Framework. Irregular, an AI security lab named in the system card, found it broadly comparable to GPT-6 Astra apart from stronger results on long multistep tasks, and neither model got past fully hardened targets in OpenAI's own challenge set, according to SiliconANGLE.

Who can use it

Approved Daybreak Red customers reach GPT-6 Cyber through OpenAI's API and Codex, and API requests must come from approved networks by default. OpenAI said the model will reach all Daybreak Red customers over the coming weeks, GPT-5.6-Cyber is opening to every customer in the tier, and critical infrastructure defenders get a Daybreak Red group of their own, SiliconANGLE reported.

On the defensive side, OpenAI is extending early Astra access through Daybreak Blue to more organizations and to individual security researchers. RuntimeWire noted that the test for defenders is whether gated exploit development produces patches faster than the same capability spreads outside controlled programs.