OpenAI has released a limited preview of its GPT-5.6 models, including Sol, Terra, and Luna, to a select group of companies in collaboration with the U.S. government. The Sol model boasts the most robust safety stack to date, with strengthened protections against high-risk activities, sensitive cyber requests, and repeated misuse. This enhanced security is designed to enable legitimate use cases such as code review, vulnerability research, and patch development, while preventing offensive activities. The model's capabilities have been demonstrated on ExploitBench, where it competed with Anthropic's Mythos Preview using only a third of the output tokens. OpenAI's goal is to provide defenders with powerful tools to identify weaknesses and strengthen systems, while enforcing strict guardrails to prevent misuse. The preview phase may involve safeguards that block or pause legitimate requests, due to the technology's dual-use nature. Evaluations have shown that GPT-5.6 Sol can produce credible memory safety leads and automate parts of real-world vulnerability research. OpenAI plans to make the models generally available in the coming weeks, following a staggered release to trusted partners approved by the government.

Source: The Hackers News