The Three Tier Release Model
OpenAI has introduced three versions of its GPT-5.6 model, named Sol, Terra, and Luna, in a limited preview restricted to a small group of companies. This staggered release is part of an ongoing collaboration with the U.S. government. Sol is the most powerful flagship model, Terra balances efficiency with capability, and Luna is optimized for speed and cost effectiveness. The company emphasizes that Sol launches with its strongest safety measures to date, including reinforced protections for sensitive cyber requests, high risk activities, and repeated misuse. OpenAI spent weeks stress testing the system against real world attacks before releasing the preview.
Cybersecurity Capabilities and Guardrails
OpenAI positions GPT-5.6 Sol as its most capable model for cybersecurity, particularly for vulnerability research and exploit development. On the ExploitBench benchmark, Sol performs competitively against Anthropic Mythos Preview while using roughly one third of the output tokens. The company intends for the model to support legitimate defensive work such as code review, patch development, debugging, security education, and vulnerability testing. Strong guardrails block offensive activities, and the system includes mechanisms to quickly remediate newly discovered jailbreaks. OpenAI acknowledges that during this preview phase, some legitimate requests may be blocked or paused for additional review due to the dual use nature of the technology.
Preview Limitations and Government Oversight
According to OpenAI’s system card, while GPT-5.6 Sol is more skilled at finding code vulnerabilities and developing exploits, it cannot autonomously execute end to end attacks against hardened targets or weaponize vulnerabilities in real attacks. Internal evaluations using the VulnLMP framework showed the model producing credible memory safety leads that could potentially lead to disclosure or control flow corruption. OpenAI suggests this indicates that substantial parts of vulnerability research are becoming automatable when models are paired with appropriate tools and infrastructure. The company plans to make all three variants generally available in the coming weeks, with the preview limited to government approved trusted partners. This release follows an executive order on AI and cybersecurity that calls for a framework to evaluate AI models with advanced cyber capabilities, as well as similar restricted releases from competitor Anthropic.
Source: The Hacker News
