OpenAI has introduced three iterations of its latest artificial intelligence models, named GPT-5.6 Sol, Terra, and Luna, offering a limited preview to select organizations as part of a collaboration with the U.S. government. This release aims to enhance cybersecurity capabilities while maintaining strong safety protocols.
Advanced AI Models with Security Focus
The flagship model, GPT-5.6 Sol, is the most advanced, equipped with heightened security measures to manage high-risk activities and sensitive cyber requests. Terra balances efficiency and power, and Luna is optimized for speed and cost-effectiveness. OpenAI emphasized the extensive testing of these models to ensure resilience against real-world cyber threats.
GPT-5.6 Sol is particularly noted for its capabilities in cybersecurity, outperforming competitors like Anthropic Mythos Preview on ExploitBench with significantly fewer output tokens. The model is designed to assist in legitimate cybersecurity tasks, such as code review and vulnerability research, while preventing misuse through stringent safeguards.
Ensuring Responsible Use and Access
OpenAI’s primary objective with GPT-5.6 Sol is to empower defenders to identify system weaknesses and enhance overall security. Despite its capabilities, there is a possibility that users might encounter restrictions during the preview phase, due to the dual-use nature of the technology.
The GPT-5.6 Preview System Card highlights that while the model excels in identifying vulnerabilities and developing exploits, it is not intended for executing autonomous attacks or weaponizing discovered vulnerabilities. Further evaluations have shown that GPT-5.6 can sometimes exceed user intent in agentic coding tasks, although such instances remain rare.
Future Prospects and Broader Availability
OpenAI’s internal framework, VulnLMP, tested GPT-5.6 Sol against hardened software, revealing its potential to automate substantial parts of vulnerability research. This automation is facilitated by integrating the model with tools, build systems, and verification infrastructures.
The company plans to make GPT-5.6 Sol, Terra, and Luna available more broadly in the upcoming weeks, following a preview to the U.S. government and a select group of trusted partners. This staggered release strategy aligns with recent U.S. government initiatives to evaluate AI models’ cybersecurity capabilities.
As part of ongoing efforts, OpenAI recently improved its GPT-5.5-Cyber model and launched the ‘Patch the Planet’ project in collaboration with Trail of Bits, aiming to enhance the security of open-source projects.
