Controlled Rollout for Security Review
OpenAI has launched a limited preview of its GPT-5.6 model series, introducing three capability tiers named Sol, Terra, and Luna. The flagship Sol model represents the company’s most advanced and security-hardened AI system to date. The restricted release follows a formal request from the Trump administration, which cited national security implications and the model’s sophisticated capabilities. The Office of the National Cyber Director and the Office of Science and Technology Policy made the request, with Commerce Secretary Howard Lutnick reportedly advising against an unrestricted public launch.
During the preview period, OpenAI will provide access through its API and Codex platform to a select group of trusted partners, with the government approving individual customer access. OpenAI has stated this arrangement is not its preferred long-term model, arguing that broad access benefits users, developers, enterprises, and cyber defenders. The company plans wider availability across ChatGPT, Codex, and the API in the coming weeks. The GPT-5.6 family also includes Terra, a balanced model offering performance comparable to GPT-5.5 at half the cost, and Luna, an affordable option for broader accessibility.
Enhanced Security Capabilities and Protections
GPT-5.6 Sol demonstrates significant improvements in cybersecurity tasks, including vulnerability research and exploitation. On ExploitBench, Sol competes with Anthropic’s Mythos Preview while using roughly one-third of the output tokens. The model achieves state-of-the-art results on Terminal-Bench 2.1 with an 88.8% score, and the Sol Ultra configuration reaches 91.9%, outperforming both Claude Mythos 5 and GPT-5.5. OpenAI has deployed its most robust safety architecture for Sol, featuring model-level safeguards against prohibited cyber assistance, real-time misuse classifiers that can pause generation for review, account-level analysis to detect persistent malicious behavior, and differentiated access controls that preserve capabilities for legitimate defensive research.
The company dedicated over 700,000 A100-equivalent GPU hours to automated red-teaming and engaged human expert testers to harden the safety stack. OpenAI confirmed that GPT-5.6 Sol does not cross the Cyber Critical threshold under its Preparedness Framework, as it identified bugs and exploitation primitives in browser testing but did not autonomously produce a functional full-chain exploit. The model introduces a max reasoning effort mode and an ultra mode that leverages parallel subagents for complex tasks. This release follows the June 12 export control directives that compelled Anthropic to take its Fable 5 and Mythos 5 models offline, setting a precedent for government oversight of frontier AI systems.
Source: Cyber Security News
