OpenAI Previews GPT-5.6 Sol, Terra, and Luna With Stronger Cybersecurity Safeguards
OpenAI released GPT-5.6 Sol, Terra, and Luna as a limited preview with restricted access and enhanced cybersecurity safeguards, in coordination with the U.S. government.
OpenAI on Friday released three versions of GPT-5.6 — named Sol, Terra, and Luna — as a limited preview to a small number of companies, marking another step in the company's ongoing engagement with the U.S. government over advanced AI capabilities. The rollout comes amid intensifying scrutiny of frontier AI models' potential for misuse in cybersecurity contexts.
Sol is the most powerful flagship model, Terra balances efficiency and capability, and Luna is optimized for speed and affordability. According to OpenAI, the models incorporate the company's most robust safety stack to date, with strengthened protections against high-risk activity, sensitive cyber requests, and repeated misuse. The company spent weeks stress-testing the systems and hardening them against real-world attacks.
GPT-5.6 Sol was also positioned as OpenAI's most capable model yet for cybersecurity work, including vulnerability research and patch development. On the internal ExploitBench benchmark, Sol competes with Anthropic's Mythos Preview using only about one-third of the output tokens. The company emphasized that the goal is to enable legitimate defensive work such as code review, vulnerability research, and security education, while maintaining strong guardrails against offensive activity and swiftly addressing newly discovered jailbreaks.
OpenAI's preview system card warns that although the model is more adept at finding vulnerabilities and developing exploits, its capabilities do not extend to autonomous, end-to-end attacks against hardened targets. Evaluations using VulnLMP, OpenAI's internal framework for testing exploit chain development, found that Sol produced credible memory safety leads, some of which could lead to disclosure, mutation, or control flow corruption. The company noted that substantial parts of real-world vulnerability research are becoming increasingly automatable when models are paired with tool use and verification infrastructure.
The staggered release follows OpenAI's rollout of an improved GPT-5.5-Cyber model to trusted defenders as part of the Daybreak initiative and the launch of the Patch the Planet project with Trail of Bits to secure open-source software. It also comes after the U.S. government permitted Anthropic to release its Mythos AI model to about 100 trusted organizations and federal agencies that operate and defend critical infrastructure, following a brief market suspension.
Earlier this month, President Donald Trump signed an executive order on AI and cybersecurity, calling for a framework that allows the federal government to evaluate AI models and designate those with advanced cyber capabilities as "covered frontier models." OpenAI intends to make GPT-5.6 Sol, Terra, and Luna generally available in the coming weeks, with the limited preview being conducted among trusted partners whose participation has been approved by the government.