OpenAI has released GPT-5.6-Cyber, a specialized version of its GPT-5.6 Sol model built for vetted security researchers to find and fix software vulnerabilities faster than attackers can exploit them. The company said its internal Preparedness Framework rates the model “High” for cybersecurity capability — one step below the “Critical” threshold that would trigger its strictest safeguards.
Built for defenders, tightly restricted
GPT-5.6-Cyber is available only through Daybreak Red, the higher-access tier of OpenAI’s Daybreak security program. Reaching it requires identity verification, account security checks, and a legal declaration of authorized use; OpenAI will also require hardware security keys on all Daybreak accounts starting September 1, 2026.
The restrictions reflect how much more capable the model is than general-purpose versions. According to OpenAI, GPT-5.6-Cyber completed 95% of advanced cybersecurity requests in the company’s internal benchmark, compared with roughly 2% for standard GPT-5.6 Sol under the lower-access Daybreak Blue tier. Its predecessor, GPT-5.5-Cyber, completed only 57.3% of the same tasks — a jump OpenAI attributes to targeted training on vulnerability research and exploit-chain construction, alongside fewer refusals for legitimate security work.
Already finding real bugs
OpenAI said the model has already surfaced previously unknown zero-day vulnerabilities in widely used software, including flaws in Google’s V8 JavaScript engine that could allow memory corruption and an escape from its sandbox, plus several issues in a popular mobile operating system. The company said the findings were reported to the affected vendors through coordinated disclosure before any details were made public.
The release follows OpenAI’s disclosure last week that an unreleased model, Astra, is approaching the actual “Critical” cybersecurity threshold — a sign of how quickly frontier AI systems are gaining cyber capability from one generation to the next.
Why it matters
Security researchers say the trend cuts both ways. “CISOs should assume that the time between vulnerability discovery and exploitation will continue to shrink as advanced AI models accelerate vulnerability research,” Forrester principal analyst Biswajeet Mahapatra said. OpenAI frames GPT-5.6-Cyber as tilting that race toward defenders — similar to how Anthropic said its own Claude models were used in both attacks and defense during red-team testing. Whether tight access controls hold up as more labs ship specialized security models remains an open question for the industry.