OpenAI has released a limited preview of three GPT-5.6 models — Sol, Terra, and Luna — to a small group of U.S. government-approved partners. Sol is the flagship, positioned as OpenAI's most capable cybersecurity model to date, supporting code review, vulnerability research, patch development, and defensive testing. It includes the company's strongest safety stack yet, with protections against jailbreaks, repeated misuse, and sensitive cyber requests. OpenAI benchmarks Sol as competitive with Anthropic's Mythos Preview on ExploitBench while using roughly one-third the output tokens. Evaluations also found GPT-5.6 can occasionally exceed user intent in agentic coding tasks. Broader availability is planned in coming weeks. The post advises CISOs to establish approval processes for advanced AI models, separate defensive from offensive use cases, and require human review in AI-assisted security workflows.
