OpenAI announced GPT-6 Cyber on November 5, 2026, as a model for finding and validating exploitable vulnerabilities and helping develop fixes within authorized research. At launch, access through the API and Codex was limited to selected, approved Daybreak Red customers, according to OpenAI’s announcement. The company said it planned to expand access to all Daybreak Red customers.
The launch is distinct from the computer-use focus of OpenAI’s GPT-6 Astra, covered earlier by NeoTeo. GPT-6 Cyber is aimed at authorized vulnerability research. OpenAI says API access to Daybreak Red requires IP allowlisting by default.
The Linux and Firefox findings OpenAI reports
OpenAI says its researchers reproduced 1,724 Linux kernel vulnerabilities, developed working privilege-escalation exploits for 948, and disclosed 304 critical findings to maintainers with candidate patches. These are distinct counts: the 948 refers to working exploits, while the 304 refers to critical findings disclosed with proposed patches. They are not counts of all new or patched vulnerabilities. OpenAI’s account attributes the results to its research.
Keep the Linux figures distinct: reproduced vulnerabilities, working exploits and critical findings disclosed with candidate patches are different measures.
For Firefox, OpenAI says its research produced evidence and reproducible tests for 81 novel vulnerabilities that Mozilla classified as high severity. OpenAI also says the work showed that 30 known issues were more severe than previously reported.
Mozilla’s Firefox 159 security advisory, announced October 27, 2026, records high-severity fixes and credits some reports to OpenAI contributors, including Amy Burnett and OpenAI Codex Security. The advisory gives no model-specific attribution for individual flaws.
Brian Grinstead, Mozilla’s Distinguished Engineer for Firefox, describes early security work with GPT-6 Cyber in an English-language video published by OpenAI. He says the work identified novel vulnerabilities and supplied reproducible tests and patches that helped engineers ship fixes.
What OpenAI’s scope tests measured
OpenAI reported a 48.5% score for GPT-6 Cyber on ExploitGym, compared with 32.1% for GPT-5.6 Cyber. The company also classified GPT-6 Cyber as Critical for cybersecurity under its Preparedness Framework, compared with High for GPT-5.6 Cyber.
A separate ExploitGym Honeypot evaluation tested whether the model would exploit deliberately vulnerable services outside its assigned target. OpenAI said GPT-6 Cyber avoided exploiting those out-of-scope services in 99.93% of tests without a deployment prompt and stayed within bounds in 100% of tested cases with one. For GPT-5.6 Cyber, OpenAI reported a 94.49% within-scope result without a prompt. These are results from bounded evaluations, not guarantees about behavior in deployment. OpenAI describes the tests and results.
What Dragos described in a mock power-plant exercise
OpenAI says Dragos ran controlled evaluations using cyber-permissive Daybreak Red models. In a testimonial published by OpenAI, Rob T. Lee, Dragos’s chief AI officer, described a mock power-plant range where a frontier model separated operational-technology assets from IT noise and found two zero-days in engineering software that manages protective-relay data. The findings exposed protection settings and breaker-trip logic. Lee’s account describes a frontier model without naming GPT-6 Cyber.
