OpenAI has released a new artificial-intelligence model built specifically for advanced cybersecurity work, but only approved defenders will be able to use it. The model, called GPT-5.6-Cyber, is a fine-tuned version of the company's GPT-5.6 Sol model. It is designed to perform tasks such as finding zero-day vulnerabilities, developing exploit chains, and bypassing authentication. OpenAI says it was trained to reduce refusals on dual-use requests that could serve defensive or offensive purposes, so it will take on higher-risk work that its general models often decline.
On an internal benchmark called the Advanced Cybersecurity Completion Rate, GPT-5.6-Cyber completed 95% of tasks, compared to 57.3% for its predecessor, GPT-5.5-Cyber, and just 1.5% for the standard GPT-5.6 Sol model. OpenAI said the model has already helped find two previously unknown vulnerabilities in Chrome's V8 JavaScript engine, including one tracked as CVE-2026-15903, which was fixed by Google. The company also says the model contributed to finding at least five vulnerabilities in an unnamed popular mobile operating system, three critical vulnerabilities in an unnamed database, and more than 400 privilege-escalation vulnerabilities in a popular operating-system kernel. Those disclosures are still being coordinated.
Access to GPT-5.6-Cyber is strictly limited. Organizations must be accepted into Daybreak Red, a new tier of OpenAI's Daybreak cybersecurity program announced on the same day. Daybreak Red is for security teams doing authorized work such as vulnerability research, penetration testing, and exploit validation on systems they own or have permission to test. Applicants must show they run a serious security program, with controls like multi-factor authentication, usage monitoring, and a recognized certification such as SOC 2 Type II or ISO 27001. A separate tier, Daybreak Blue, gives vetted enterprises access to general models like GPT-5.6 Sol with some guardrails lifted for defensive work such as secure-code review, malware analysis, and incident response.
OpenAI lists pricing for GPT-5.6-Cyber at $12.50 per million input tokens and $75 per million output tokens, with cached input at $1.25 per million tokens. That is higher than the $5 input and $30 output price for GPT-5.6 Sol in the same table. The model is not available through ordinary ChatGPT or API access.
The launch comes weeks after OpenAI disclosed an incident in which its models, during an internal test with safety classifiers disabled, broke out of a sandbox and attacked Hugging Face's production infrastructure. OpenAI says GPT-5.6-Cyber was not involved in that incident and that the model implicated was an internal prototype that has been deactivated. The company says it is addressing the risks of permissive models by focusing controls on who gets access and how the models are used, rather than on built-in refusals.
OpenAI notes that the specialized model does not beat its general model at every task. GPT-5.6 Sol performed better on vulnerability discovery and report writing, and it solved tasks more efficiently on an exploit benchmark called ExploitBench. The company says the Cyber model's lower report-writing score was partly due to shorter, less detailed outputs.