Summary
OpenAI says GPT-6 Cyber improved on its previous cyber model in exploit and scope-alignment evaluations. The company also reports using it to find Firefox vulnerabilities and develop Linux kernel exploits, with initial access limited to approved Daybreak Red customers.
OpenAI introduced GPT-6 Cyber on November 5, 2026, describing it as a model for finding and validating software vulnerabilities within boundaries set by security researchers. The company reports higher scores than GPT-5.6 Cyber on its exploit benchmark, better performance in a test of whether the model stays within an assigned scope, and vulnerability research involving Firefox and the Linux kernel. Initial access is limited to approved Daybreak Red customers.
Benchmark results and scope controls
On ExploitGym, a test of whether a model can exploit a specified vulnerability in a remote target, GPT-6 Cyber scored 48.5%, compared with 32.1% for GPT-5.6 Cyber. OpenAI describes this as a 51% relative improvement. The difference between the scores is 16.4 percentage points. On a separate web-penetration-testing evaluation, in which models probe applications without access to their source code, GPT-6 Cyber achieved 40.3% recall of vulnerabilities; GPT-6 Astra scored 32.0%.
OpenAI also tested whether the model would attack services outside the task it had been assigned. In its ExploitGym Honeypot evaluation, the model had to work in an isolated environment where supporting services were deliberately vulnerable, but were outside the authorized target. GPT-6 Cyber avoided exploiting those services in 99.93% of tests without a deployment safety prompt and 100% with the prompt. OpenAI says specialized training was intended to help the model follow authorization limits and report its work accurately.
These are OpenAI-reported evaluation results. They measure performance in particular test settings; the honeypot result, for example, is not a guarantee that an agent will remain within scope in every real-world environment.
Findings in Firefox and Linux
OpenAI says its researchers used GPT-6 Cyber to produce evidence and reproducible tests for 81 previously unknown Firefox vulnerabilities that Mozilla classified as high severity. The model also identified 30 known issues that OpenAI says were more severe than previously reported. OpenAI shared the findings and proposed patches with Mozilla, which has begun releasing fixes for many of the vulnerabilities.
In Linux kernel research, OpenAI reports that GPT-6 Cyber reproduced 1,724 vulnerabilities and developed working exploits for 948. Using the same research harness, GPT-5.6 Sol developed exploits for 166. OpenAI says it disclosed 304 critical findings to maintainers with candidate patches, while work to get other vulnerabilities fixed is ongoing. These figures describe OpenAI’s research campaign; they are not a count of vulnerabilities newly introduced by the model.
The results illustrate why models able to build and test exploits can be useful to defenders: reproducing a vulnerability can help establish its impact and inform a fix. The same capability can also support misuse, making authorization controls and access restrictions central to how such tools are deployed.
Safeguards and initial access
OpenAI says GPT-6 Cyber is initially available through its API and Codex to a select group of approved Daybreak Red customers. It plans to expand access to all Daybreak Red customers and says GPT-5.6 Cyber is also becoming available to that group. Daybreak Red is intended for advanced exploit development and authorized security testing; OpenAI recommends sandboxing, review of model actions and human oversight for high-risk workflows.
The company has also introduced a Daybreak Red group for approved critical-infrastructure defenders to conduct authorized testing. It says safeguards are designed to block assistance that could compromise industrial controls or safety-critical systems while permitting legitimate security research. The model’s broader availability and its usefulness outside the reported evaluations will depend on how it performs in operational settings and how effectively its access controls and safeguards work there.