Dream's Hero scores 96.6% on CyberGym

Dream says its Hero system, powered by Hercules, scored 96.6% on UC Berkeley’s CyberGym Level 1, passing 1,456 of 1,507 tasks—the highest score it reports on the benchmark. Dream’s benchmark write-up says the result comes from an on-premises model-and-harness system with no external API calls, not a general-purpose model in isolation. Defenders should file it as evidence that specialized, sovereign systems can beat frontier-model headlines on narrow cyber work, not as a reason to expose production code to autonomous tools; OpenAI’s Critical-tier Astra rollout shows why capability claims now need deployment gates.