OpenAI previews GPT-5.6 Sol, Terra, Luna for vulnerability research
OpenAI announced three GPT-5.6 variants focused on cybersecurity, with Sol positioned as the flagship for vulnerability analysis and exploit development. All three will be publicly available within weeks.
OpenAI this week previewed GPT-5.6 in three configurations — Sol, Terra, and Luna — all aimed at cybersecurity workflows. The company plans to make all three publicly available within weeks. GPT-5.6 Sol is described as the most capable variant for long-horizon security tasks, including vulnerability research and exploitation.
The models ship with model-level refusals for prohibited cyber assistance, including attempts to disguise intent or jailbreak the system. OpenAI frames these guardrails as a first boundary layer for what the model will and won't help with, though the company has not disclosed whether Terra and Luna carry the same refusal tuning or are differentiated by capability tier.
What stands out
- 01Three-tier release. Sol is the flagship for security research; Terra and Luna's roles remain unspecified. The naming convention suggests a hierarchy, but OpenAI has not published parameter counts, context lengths, or benchmark comparisons.
- 02Explicit exploit capability. OpenAI positions Sol for "vulnerability research and exploitation" — a rare acknowledgment that the model is designed to assist in offensive security tasks, not just defensive analysis.
- 03Model-level refusals. The refusal layer is trained into the weights, not just an API filter. This approach is harder to bypass than prompt-level guardrails but also harder to ablate for researchers who need unrestricted access.
- 04No release date precision. OpenAI has not announced a specific launch window beyond "the coming weeks." No pricing, API tier, or local-weight availability has been disclosed.
- 05
