OpenAI, GPT-5.6-Cyber announced its model on August 10, offering a special option for advanced cybersecurity research. The company trained the GPT-5.6 Sol-based model to find zero-day vulnerabilities, develop exploit chains, and authoritative security testing. However, OpenAI does not open access to everyone, only users who have completed authentication and Daybreak Red Researchers and organizations that receive approval can use the model. Thus, GPT-5.6-Cyber will focus on a different usage area than the general-purpose models offered in ChatGPT.
The word “Cyber” in the model name directly describes this cyber security focus. OpenAI ensures that GPT-5.6-Cyber yields fewer rejections than the general-purpose model for certain dual-use security tasks. In the company’s Advanced Cybersecurity Completion Rate test, the model performed advanced tasks. 95 percent completed. In the same test, GPT-5.5-Cyber got 57.3 percent, GPT-5.6 Sol got 1.5 percent, and the Daybreak Blue version got 2 percent.
On the technical side 400k token context window And 128 thousand tokens maximum output It has support. In addition, the model processes text and visual inputs and uses web search, file search and code interpreter tools via the Responses API. “Hosted shell”, “computer use” and MCP support also allows developers to combine different tools in security studies in the same process. OpenAI for 1 million login tokens $12.50for 1 million exit tokens 75 dollars He wants a fee.
Performance tests also show that the model does not produce the same results in every task. GPT-5.6-Cyber beat GPT-5.6 Sol and GPT-5.5-Cyber in the ExploitGym test, which measures the production of exploits working from known vulnerabilities. Additionally, the Cyber version scored higher than GPT-5.6 Sol in OpenAI’s internal testing that measures zero-day vulnerabilities. On the other hand, in the Vulnerability Discovery and Report Writing test, GPT-5.6 Sol achieved higher results by preparing more detailed reports.
GPT-5.6-Cyber finds vulnerabilities in real software
OpenAI has found two new vulnerabilities that can be chained in GPT-5.6-Cyber and the V8 JavaScript engine that Chrome uses. Google one of these CVE-2026-15903 code and closed the relevant vulnerability. The company also identified at least five critical vulnerabilities in a popular mobile operating system and three in a widespread database. which may lead to privilege escalation in an operating system kernel. More than 400 problems found, but OpenAI did not share the names of related products.
The overall GPT-5.6 family also showed higher results in cybersecurity tests than the previous generation. GPT-5.6 achieved 73.5 percent on ExploitBench, while GPT-5.5 achieved 47.9 percent with the same token budget. At ExploitGym, GPT-5.6 achieved a 24.9 percent success rate in two hours and 33.7 percent in six hours. GPT-5.5 reached a maximum level of 15.1 percent in the same test.
OpenAI offers GPT-5.6 Sol access via Daybreak Blue to most defense teams for advanced work Daybreak Red is using. This scope includes exploit verification, authorized penetration testing, red teaming and advanced security research. The company controls access through authentication, account security, usage tracking and separate terms of use. As part of the Preparedness Framework, OpenAI strengthens the cybersecurity capabilities of GPT-5.6 Sol and GPT-5.6-Cyber. high level and explains that neither model reaches the Critical threshold.
Join Channel