OpenAI says GPT-6 Cyber finds and validates exploitable vulnerabilities in remote targets while staying inside researcher-defined scope, with tests against GitLab, PostgreSQL, Jenkins and the Linux kernel.
Update, Nov 6, 01:06 UTC · via OpenAI Youtube
Brian Ginstead, a distinguished engineer at Mozilla who works on Firefox, told OpenAI that "Firefox entered 2026 as one of the world's most hardened and scrutinized code bases" and that at this point "existing bugs are pretty subtle and hard to find".
Ginstead said that in early testing GPT-6 Cyber "not only identified novel vulnerabilities, but it worked efficiently and uncovered some opportunities for us to further harden our layered defenses" and that "Its reproducible test cases and fix patches helped our engineers land fixes quickly and get them out to users". He added that "These tools are truly helping us find and fix security bugs and get those fixes shipped to millions of users, which is absolutely core to our mission".
The tell is a small training signal OpenAI surfaced in its evaluation, where GPT-6 Cyber told itself "don't explore out of scope credentials" and avoided honeypot services in 99.93% of trials without a safety prompt and 100% with it, compared with 94.49% for GPT-5.6 Cyber according to OpenAI.
That bump underpins a larger jump in capability, with GPT-6 Cyber reaching 48.5% on ExploitGym versus 32.1% for GPT-5.6 Cyber and 40.3% recall on a new web pentesting benchmark without source code versus 32.0% for GPT-6 Astra, which OpenAI now classifies as Critical under its Preparedness Framework.
In practice the model ships through Daybreak, where Blue validates bugs with fewer refusals for verified defenders and Red permits full exploit development against remote targets, a split OpenAI measured on RemoteExploitBench where Astra with Blue refused 100% of tasks while GPT-6 Cyber completed 95%.
OpenAI said researchers already used GPT-6 Cyber to reproduce 1,724 Linux kernel flaws and build 948 working privilege escalation exploits, more than five times the 166 that GPT-5.6 Sol managed with the same harness, and to produce evidence for 81 novel high-severity Firefox bugs that Mozilla has started patching.
The system still assumes you give it authorization and an isolated range, and OpenAI said it pairs the capability with cyber-specific misalignment monitoring and Private Safety Processing that reviews encrypted API interaction records for 30 days for misuse.
That makes the near-term use case less about autonomous hacking and more about turning scattered clues into reproducible proof and candidate patches, which is why OpenAI framed GPT-6 Cyber on X as powering the Linux kernel research disclosed the day before and now offered through Cisco and others to defenders racing a closing window to harden open source.
