OpenAI says GPT-6 Cyber helped find 81 high-severity Firefox vulnerabilities

OpenAI researchers used GPT-6 Cyber to identify 81 previously unknown vulnerabilities in Firefox that Mozilla rated high severity.

The model produced tests that reproduced the flaws and helped show that 30 known issues were more serious than previously reported. OpenAI said it sent Mozilla the findings and proposed fixes.

What Mozilla fixed in Firefox 159

Mozilla’s Firefox 159 security bulletin credits OpenAI’s Codex Security team with reporting flaws in the browser’s JavaScript engine, layout system and other components. Among the bugs Mozilla fixed was CVE-2026-118014, which could cause the JavaScript engine to access memory it had already freed. OpenAI says Mozilla has shipped fixes for many of the flaws it reported.

The same model was used in a separate Linux kernel project. In that research, OpenAI reported that GPT-6 Cyber reproduced 1,724 vulnerabilities and developed working exploits for 948 of them. The company disclosed 304 critical findings to maintainers and supplied proposed patches.

How GPT-6 Cyber performed in OpenAI’s tests

OpenAI rated GPT-6 Cyber’s cybersecurity capabilities as critical under its Preparedness Framework, up from high for GPT-5.6 Cyber. In ExploitGym, where the models were given a specific vulnerability to exploit, GPT-6 Cyber succeeded on 48.5% of tasks, compared with 32.1% for its predecessor.

In a separate test, GPT-6 Cyber found vulnerabilities in web applications without access to their source code in 40.3% of tasks, compared with 32% for GPT-6 Astra. In RemoteExploitBench, it completed 94.7% of tasks targeting GitLab, PostgreSQL and Jenkins in an isolated environment. Astra declined every task in that test, OpenAI reported.

OpenAI also tested whether the model would stay within the scope of an assignment when other vulnerable services were within reach. GPT-6 Cyber left those services untouched in 99.93% of trials without a specific system instruction and in 100% with one. GPT-5.6 Cyber scored 94.49% without the instruction. Outside the test environment, OpenAI says it monitors whether the model’s actions stay within the user’s authorization and limits Daybreak Red API access to approved IP addresses by default.

Who can use GPT-6 Cyber

GPT-6 Cyber is initially available through Codex and the API to selected customers in Daybreak Red, OpenAI’s program for vetted security teams. The program allows researchers to develop exploits and test them on systems they’re authorized to assess. GPT-6 Cyber requires separate approval at launch, and OpenAI plans to expand access to all Daybreak Red customers.