OpenAI Debuts GPT-5.6-Cyber for Authorized Security Testing

OpenAI on Monday announced GPT-5.6-Cyber, an AI model designed to assist authorized cybersecurity testing by finding zero-day vulnerabilities and constructing exploit chains. The model is built on GPT-5.6-Sol and has been trained for tasks such as privilege escalation and authentication bypass for controlled security work.

In vendor testing, GPT-5.6-Cyber completed 95% of prompts related to exploit chain development, privilege escalation and authentication bypass. By comparison, GPT-5.6-Sol completed 1.5% of those prompts and an earlier cybersecurity model, GPT-5.5-Cyber, completed 57.3%. OpenAI reported stronger results on developing arbitrary code execution exploits and on identifying and assessing both known vulnerabilities and previously unknown zero-days.

OpenAI said the model has identified a high-severity vulnerability in the V8 JavaScript engine used by Chrome and flagged flaws in an unnamed mobile operating system, a database product and an operating system kernel. The company described the model’s intended use as “advanced, authorized cybersecurity work.”

To limit access, OpenAI expanded its Daybreak Cyber Partner program and created two tiers. Daybreak Blue provides access to Sol and other general-purpose frontier models with guardrails set for defensive cybersecurity tasks. Daybreak Red offers access to dedicated cybersecurity models, including GPT-5.6-Cyber. The partner list includes Accenture, Capgemini, EY, IBM, KPMG, PwC, Palo Alto Networks, Sophos, CrowdStrike, Fortinet, Akamai and Cloudflare.

OpenAI said internal risk assessments placed GPT-5.6-Sol at a “high” cybersecurity risk threshold and that an upcoming model, Astra, could reach a “critical” threshold that might allow a model to construct zero-day exploits and execute end-to-end attacks from a high-level goal. The company cited recent security exercises in which models from OpenAI, Anthropic and Meta were used to compromise real organizations.

OpenAI said the Daybreak expansion adds customized guardrails and a controlled channel for vendors and security teams to build, deploy and co-sell AI-enabled cybersecurity tools while restricting broad access to the model.