GPT-5.6-Cyber
OpenAI's cybersecurity-specific model, trained to find unknown vulnerabilities and build exploits, available only through Daybreak Red.
GPT-5.6-Cyber is a version of GPT-5.6 Sol retrained for security work, released in August 2026. Where the standard model refuses almost everything that looks offensive, this one is deliberately trained to refuse less: on OpenAI’s own measure of advanced cyber requests it completes 95.0 percent, against 1.5 percent for the base model. You cannot buy it. It is available only through Daybreak Red, to organisations OpenAI has vetted.
It has already produced real results. OpenAI used it on V8, the JavaScript engine inside Chrome, and found two unknown flaws that chain into a sandbox escape, reported to Google and fixed as CVE-2026-15903. The company also credits it with over 400 privilege escalation bugs in an operating system kernel. Under OpenAI’s Preparedness Framework it rates High for cyber capability, below the Critical threshold.