OpenAI Previews GPT-5.6 Sol With Restricted Access and Stronger Cyber Safeguards

OpenAI has released three versions of GPT-5.6—Sol, Terra, and Luna—as a limited preview to a small number of companies in coordination with the U.S. government. Sol is the flagship model with the most robust safety stack to date, designed for vulnerability research and exploitation while enforcing strong guardrails against offensive activity. Terra balances efficiency and power, while Luna is optimized for speed and affordability.

According to OpenAI, GPT-5.6 Sol is competitive with Anthropic Mythos Preview on ExploitBench using only one-third of the output tokens. The model aims to enable legitimate cybersecurity work such as code review, vulnerability research, patch development, debugging, and security education, while blocking prohibited cyber assistance and swiftly remediating jailbreaks. OpenAI warns that during the preview, users may encounter safeguards that block legitimate requests or require additional review due to the dual-use nature of the technology.

The GPT-5.6 Preview System Card indicates that while the model is adept at finding vulnerabilities and developing exploits, it does not extend to autonomous end-to-end attacks against hardened targets. Evaluations using VulnLMP found that GPT-5.6 Sol produces credible memory safety leads that could lead to disclosure, mutation, or control flow corruption. OpenAI plans to make the models generally available in the coming weeks, following previews to the U.S. government and a limited group of trusted partners.

This release follows an executive order by President Donald Trump on AI and cybersecurity, and comes after Anthropic was permitted to release its Mythos AI model to trusted companies and federal agencies. OpenAI also released an improved GPT-5.5-Cyber model and launched the Patch the Planet project with Trail of Bits to secure open-source projects.