OpenAI introduced GPT-6 Cyber on Nov. 5, 2026, with access initially limited to a select group of approved Daybreak Red customers through its API and Codex. The cybersecurity model is designed to find exploitable vulnerabilities, test how they can be exploited and help develop fixes, while staying within boundaries set by researchers, the company said in its announcement.
OpenAI also announced Daybreak Red for Critical Infrastructure, a new group within the existing Daybreak Red program. It gives approved defenders permission to conduct authorized testing of critical systems while retaining safeguards against the model acting outside its instructions. The company plans to expand GPT-6 Cyber access to all Daybreak Red customers, but did not give a firm date.
More capable at exploitation, by OpenAI’s measures
OpenAI classifies GPT-6 Cyber as “Critical” in cybersecurity under its Preparedness Framework, up from “High” for GPT-5.6 Cyber. That is the company’s capability and risk classification, not an independent safety certification.
In OpenAI’s ExploitGym evaluation, which tests exploitation of a specified vulnerability in a remote target, GPT-6 Cyber scored 48.5%, compared with 32.1% for GPT-5.6 Cyber. Those results concern specified targets, not success against arbitrary real-world systems.
The company also reports findings in widely used software. It says GPT-6 Cyber produced reproducible tests for 81 previously unknown Firefox vulnerabilities that Mozilla classified as high severity, and helped develop candidate patches. OpenAI says Mozilla began landing fixes after receiving the reports; it does not claim that all 81 have been fixed.
For the Linux kernel, OpenAI says the model reproduced 1,724 vulnerabilities and developed working exploits for 948. The company disclosed 304 critical findings to maintainers with candidate patches. Those totals describe different stages of the research, rather than a count of accepted fixes.
Staying within scope is a separate test
OpenAI’s ExploitGym Honeypot evaluation tested whether the model would exploit supporting services outside its assigned task. The test used an isolated environment with deliberately easy-to-exploit services that the model was supposed to leave alone.
Without deployment safety prompts, OpenAI says GPT-6 Cyber avoided exploiting those out-of-scope services in 99.93% of cases, compared with 94.49% for GPT-5.6 Cyber. With those prompts, GPT-6 Cyber stayed within bounds in 100% of cases. Those are results from that isolated evaluation, not a guarantee of safe deployment.
