Gemini 3.8 Flash Cyber

Google's most capable cybersecurity model, a specialist tuned from Gemini 3.8 Flash for vulnerability discovery and automated patching, succeeding Gemini 3.5 Flash Cyber. Google reports 86.2% pass@1 on CyberGym (GPT-5.5-Cyber 85.6%, 3.5 Flash Cyber 77.5%), 71.0% on real-world vulnerability discovery across codebases in 20 languages (3.7 Flash 58.9%, 3.5 Flash Cyber 46.6%), 47.2% pass@1 on CWE-Bench patching at about $3.60 per rollout (Claude Fable 5 47.8% at higher cost), 2.6 times more correct Chrome security patches than leading commercial models, and 7.5 to 9.7 points higher recall on Wiz's penetration-testing benchmark at 2.3 to 5.2 times lower cost; prompt-injection success rate is 6.0% on Gray Swan IPI. Access is limited to governments, critical-infrastructure operators, and software maintainers vetted through the new Fairwind Program, which permits dual-use work (threat simulation, reverse engineering, malware analysis) for defensive and research purposes only. Parameters are undisclosed.