Entity · DiggingBeagle record
GPT-6 Cyber
Restricted cyber-specialized GPT-6 model; Critical capability classification and authorization-sensitive distribution.
Record description
GPT-6 Cyber is OpenAI's cyber-specialized variant of GPT-6 Astra, introduced November 5, 2026 for approved vulnerability discovery, exploitation validation, defensive code review and remediation research. OpenAI classifies its cybersecurity capabilities as Critical under its Preparedness Framework. Daybreak Red access is restricted and requires user and organization verification with scope controls; an approved researcher is not thereby authorized to probe arbitrary systems. The model is neither a generally distributed consumer offering nor downloadable open weights. The vendor reports stronger exploit-benchmark performance than GPT-5.6 Cyber, but individual benchmark scores describe particular harnesses and policies, not live attack rates.
Evaluation results differ sharply by protocol. In a controlled scope test reported by OpenAI, the model avoided exploiting easy out-of-scope supporting services in 99.93% of trials without its deployment prompt and 100% with that prompt, compared with 94.49% for GPT-5.6 Cyber without the prompt. The same November 5 system card reports a negative external comparison: FrontierCyber 90/288 for GPT-6 Cyber versus 96/288 for Astra on the updated task set, with neither solving the 15 highest-difficulty challenges. The controlled scope results do not certify all downstream tool executors, and the independent benchmark prevents interpreting specialized cyber training as improvement on every task. The Linux exploit pipeline and Mozilla's Firefox remediation trail provide separate real-software evidence under their own research protocols.
Deployment identity matters. GPT-6 Cyber's November 5 initial availability is to selected Daybreak users through the specified API and Codex routes. A separate November 4 GovCloud announcement concerns GPT-5.6 Cyber through Amazon Bedrock in AWS GovCloud (US-East), where OpenAI approval and AWS provisioning are different requirements. Neither offering allows unrestricted scanning of systems a user does not control. A case report should identify model version, access route, actor authorization and executor tools, rather than using the broad Daybreak brand as proof of a particular deployment.