OpenAI has introduced GPT-6 Cyber, a specialised variant of GPT-6 Astra designed for authorised vulnerability discovery, exploit development and web penetration testing. The company has classified the model as “Critical” for cybersecurity capability under its Preparedness Framework, matching the designation assigned to GPT-6 Astra while offering cybersecurity-specific post-training.
Access is being restricted through the Daybreak Red programme. Participating organisations must complete identity verification and meet security requirements that include single sign-on, multifactor authentication and API IP allow-listing. The deployment structure is intended to make the model available to selected defensive-security organisations while limiting its availability for unsupervised or unauthorised use.
GPT-6 Cyber scored 48.5% on ExploitGym, compared with 32.1% for GPT-5.6 Cyber, representing a 51% relative improvement. OpenAI also tested whether the model stayed within the boundaries of assigned security work. It avoided exploiting out-of-scope services in 99.93% of tests without the deployment prompt and in all tests when that prompt was applied.
The company said the model has been used to identify and validate vulnerabilities in security-critical codebases including Linux and Firefox. A separate pipeline run reproduced 1,724 vulnerabilities in about 30 hours and generated proofs of concept for hundreds of information-disclosure and local privilege-escalation issues. These figures describe pipeline outputs and candidate issues; they should not be interpreted as 948 patches already accepted by Linux maintainers.
The release arrives as Indian banks, NBFCs, fintech companies and government agencies confront increasingly automated attack campaigns. It also raises procurement and governance questions for cybersecurity providers that want to use frontier models in vulnerability research. The access controls make organisational security, auditability and clearly defined authorisation part of the product rather than external operating procedures.
GPT-6 Cyber’s system card also reports a 47.2% standard score on ProtocolQA, rising to 50.9% when refusals and safe completions are counted as correct. Both scores remained below the benchmark’s indicative High threshold of 54%. The model is therefore being presented as a substantial capability increase with significant operational restrictions, rather than as a generally available security assistant.
