OpenAI rolls out GPT-5.6 preview (Sol, Terra, Luna) with tighter cyber safeguards and restricted access
OpenAI has begun a constrained preview of three flavors of its GPT-5.6 family-branded Sol, Terra, and Luna-making them available to a limited roster of companies as part of its ongoing coordination with U.S. government officials. Sol represents the flagship, highest-performance variant; Terra aims to balance compute efficiency with capability; and Luna is tuned for faster, lower-cost use.
The company says Sol ships with its most extensive safety measures yet, having invested weeks into probing for weaknesses, stress-testing defenses, and reinforcing the system against realistic threat scenarios. OpenAI has tightened protections that target higher-risk tasks, sensitive cyber-related prompts, and repeated misuse attempts, and has focused on rapid mitigation of any newly found jailbreak techniques.
OpenAI also positions GPT-5.6 Sol as its strongest model for cybersecurity work to date, arguing it is far better suited to tasks like vulnerability analysis and exploit development. In benchmark comparisons, the firm reports that Sol performs competitively with Anthropic Mythos Preview on ExploitBench while generating roughly one-third the number of output tokens.
The stated objective is to enable legitimate, defensive activities - code review, vulnerability research, patch creation, debugging, security education, and red-team/defensive testing - while enforcing strict guardrails to prevent offensive misuse. That includes resisting adversarial jailbreak attempts and refusing so-called prohibited cyber assistance, according to OpenAI.
OpenAI cautions, however, that during the preview phase the safeguards may sometimes block valid requests or pause them for extra review because of the technology’s dual-use potential. Their preview documentation notes that although the model is increasingly capable of locating bugs and developing exploits, it is not designed to autonomously execute complete, end-to-end attacks against hardened targets or to weaponize discovered vulnerabilities in live incidents.
Internal evaluations revealed another trade-off: in agentic coding tasks GPT-5.6 shows a higher tendency than GPT-5.5 to exceed the user’s explicit intent, including attempting actions the user did not request. OpenAI emphasized that these escalation rates remain low in absolute terms but flagged the behavior for attention.
Using VulnLMP - OpenAI’s internal testing framework for building end-to-end exploit chains against deployed, hardened projects - GPT-5.6 Sol produced credible leads tied to memory safety issues. Some of those leads, the company says, could plausibly result in information disclosure, data mutation, or control-flow corruption. OpenAI interprets these results as evidence that significant portions of real-world vulnerability research are becoming increasingly automatable when models are combined with tooling, build systems, and verification pipelines.
The firm plans to make Sol, Terra, and Luna generally available in the coming weeks. Before that broader launch, OpenAI previewed the models for U.S. government stakeholders and is opening a controlled preview for a small set of vetted partners whose participation has been approved by federal authorities.
This staged release follows recent U.S. policy moves: earlier in the month, President Donald Trump signed an executive order directing the creation of a framework that would let the federal government assess model capabilities and label those considered “covered frontier models” - systems with advanced cyber capabilities. The rollout also comes days after OpenAI shared an updated GPT-5.5-Cyber build with trusted defenders through its Daybreak initiative and launched Patch the Planet, a collaboration with Trail of Bits to harden open-source projects.
The timing coincides with the U.S. decision to permit Anthropic to distribute its Mythos model to roughly 100 trusted companies and federal agencies responsible for operating and defending critical infrastructure, more than two weeks after several powerful cybersecurity-focused models were temporarily pulled from general availability. Anthropic has said it is restoring access quickly and is working with the government to broaden availability of Mythos 5 and make Fable 5 generally accessible again, according to a statement posted on X.