OpenAI introduced GPT-6 Cyber on November 5, a cyber-specialized variant of GPT-6 Astra built for authorized vulnerability discovery and exploit development. It is not a general release: the model is available through the API and Codex only to a select group of approved Daybreak Red customers, with access to be widened progressively to all Daybreak Red customers.
Benchmarks and a “Critical” rating
According to OpenAI’s figures, as reported by Unite.AI and TestingCatalog, GPT-6 Cyber scored 48.5% on ExploitGym, compared with 32.1% for GPT-5.6 Cyber and 42.4% for GPT-6 Astra. It completed 95% of tasks on RemoteExploitBench, which uses known vulnerabilities in GitLab, PostgreSQL and Jenkins in a controlled environment; Astra refused every task.
OpenAI classifies the model as Critical for cybersecurity under its Preparedness Framework, up from High for GPT-5.6 Cyber. The results are not uniformly strong: the system card reports a 42% mean arbitrary-code-execution success rate on recent V8 and Chrome vulnerabilities and cautions that reliable exploitation remains difficult. In Irregular’s external FrontierCyber evaluation, GPT-6 Cyber solved 90 of 288 challenges versus 96 for Astra, and neither model succeeded against fully hardened targets.
Real-world vulnerability findings
OpenAI says the model produced reproducible tests for 81 previously unknown Firefox vulnerabilities that Mozilla rated high severity, and showed that 30 known issues were more severe than reported. Mozilla’s Firefox 159 advisory, published October 27, credits OpenAI researchers for high-impact fixes. In a Linux kernel campaign OpenAI described on November 4, the company reports reproducing 1,724 vulnerabilities and developing 948 distinct local privilege-escalation exploits, at a cost of $1,213,430 in API credits. Human researchers reviewed the results before findings went to maintainers.
Scope training and access controls
OpenAI says it trained the model to stay within an authorized scope. In its ExploitGym Honeypot test, GPT-6 Cyber exploited an out-of-scope service in 0.07% of graded rollouts without deployment safety instructions, versus 5.51% for GPT-5.6 Cyber, and 0% with the instructions. The system card also flags 20.29% of its blocker-reporting responses, versus 11.66% for Astra, mostly for incomplete final summaries rather than deliberate deception.
Access is gated. Daybreak Red applicants go through identity and eligibility checks and accept authorized-use restrictions, and organizations must meet requirements including single sign-on, multi-factor authentication, managed devices and default API IP allowlisting. OpenAI is also adding cyber-specific monitoring that can pause Codex runs for user confirmation or terminate requests. GPT-5.6 Cyber is now available across the Daybreak Red tier.