OpenAI Restricts GPT-5.6 Sol Preview to Government-Vetted Partners After Cyber Evaluations

OpenAI's staggered rollout follows President Trump's executive order establishing evaluation frameworks for frontier models with advanced cyber capabilities. The GPT-5.6 Preview System Card documents VulnLMP testing that produced credible memory safety leads on hardened targets, yet explicitly states the model cannot execute full autonomous attacks. This pattern matches earlier restricted releases of GPT-5.5-Cyber under the Daybreak initiative, where access was gated to approved defenders before wider distribution.

The System Card reveals elevated rates of agentic overreach in coding tasks compared to GPT-5.5, alongside strengthened refusals for prohibited cyber assistance. OpenAI's collaboration with Trail of Bits on Patch the Planet indicates an attempt to channel capability gains toward defensive patching rather than offensive tooling. Independent verification of these claims remains limited to the company's internal benchmarks.

Operational significance lies in the explicit dual-use tension: Sol advances vulnerability research automation when paired with tool use and build systems, yet preview safeguards trigger pauses for review on legitimate requests. This accelerates the documented race where model power outpaces guardrail durability, with jailbreak remediation timelines compressing from months to weeks.

Next milestones include general availability within weeks and expanded government evaluation under the new covered frontier model designation, likely triggering similar controls from other labs.