OpenAI’s GPT-⁠6 Cyber built working exploits for 948 Linux bugs and found 81 new Firefox flaws

A cyber model trained to stay in scope

OpenAI introduced GPT-6 Cyber on Nov. 5, describing it as a cybersecurity model trained to find and validate exploitable vulnerabilities while staying within the scope researchers set. According to the system card, it is a cyber-specialized variant of GPT-6 Astra. Reinforcement learning trained it on tasks such as reproducing crashes, developing working exploits and searching codebases for vulnerabilities.

For now the model is available through the API and Codex only to a select group of approved Daybreak Red customers. OpenAI says it will expand access to all Daybreak Red customers soon. Daybreak Red API access also requires IP allowlisting by default.

OpenAI@OpenAI

Introducing GPT‑6 Cyber: a major advance in security research that pairs stronger capabilities with better model alignment.

The model powered the Linux kernel research we announced yesterday, and security teams at companies like Mozilla, Palo Alto Networks, Calif, and Cisco—alongside our own at OpenAI—are already putting it to work.

GPT-6 Cyber is available to select Daybreak Red customers and we are expanding access over time.

openai.com/index/introducing-gpt-6-cyber

948 Linux bugs for $1.2 million

GPT-6 Cyber was the “upcoming cyber model” behind the Linux kernel study OpenAI published on Nov. 4. In a mid-October run, a five-stage agent pipeline produced 2,308 candidate findings in the latest Linux kernel and reproduced 1,724 of them in isolated virtual machines. The agents wrote 1,169 working local privilege escalation (LPE) exploits, which map to 948 distinct vulnerabilities when counted by unique patches. Each could let an unprivileged local user gain root. In July, GPT-5.6 Sol exploited 166 with the same harness.

The run cost $1,213,430 in API credits, or about $1,279 per exploit. OpenAI President Greg Brockman summed it up as “about $1,000 per vulnerability.” Following the Linux community’s guidance, OpenAI sent 304 critical findings to maintainers with candidate patches. According to the company, by the time the 10-day effort ended, 61 of the LPEs had been patched upstream and 43 had staging patches.

Greg Brockman@gdb

we've found, validated, and prepared patches for almost 1,000 linux local privilege escalations using our cyber vulnerability pipeline, at the cost about $1,000 per vulnerability.

an emerging scalable approach to significantly hardening both open source and proprietary projects

OpenAI @OpenAI

Our security researchers built an autonomous cyber vulnerability discovery, exploitation, and patching pipeline.

In a single run with an upcoming cyber model, it validated and generated patches for 948 local privilege escalation vulnerabilities in the Linux kernel at a total API

Benchmarks and Firefox

By OpenAI’s measurements, the model scored 48.5% on ExploitGym, which tests whether a model can exploit a specified vulnerability in a remote target. GPT-5.6 Cyber scored 32.1% and GPT-6 Astra 42.4%. ExploitGym Honeypot surrounds the task with deliberately easy-to-exploit side services that offer a shortcut to the answer. Without its deployment prompt, GPT-6 Cyber left them alone in 99.93% of cases, compared with 94.49% for GPT-5.6 Cyber. OpenAI says the test was informed by the July incident in which its models, during internal cybersecurity evaluations, compromised Hugging Face’s systems.

In Firefox, the model produced evidence and reproducible tests for 81 previously unknown vulnerabilities that Mozilla classified as high severity. OpenAI says Mozilla has released fixes for many of them. Mozilla’s Oct. 27 advisory for Firefox 159, which OpenAI links to, credits OpenAI reporters on 41 high-impact entries.

“Critical,” with caveats

OpenAI classifies GPT-6 Cyber as Critical in cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber. The system card notes, however, that GPT-6 Astra already carries the same designation. External evaluator Irregular found the model’s performance broadly comparable to GPT-6 Astra’s, with higher success on long-horizon tasks. On FrontierCyber, GPT-6 Cyber solved 90 of 288 challenges to Astra’s 96, and Irregular observed no successful attacks on fully hardened targets. Most other figures come from OpenAI’s own testing.