Google DeepMind has introduced Gemini 3.8 Flash Cyber, a model built for defenders who need to find and fix software vulnerabilities quickly, available to a group of trusted defenders through the company’s Fairwind Program as an early-access security release. DeepMind says the model is designed to search complex codebases, identify hidden weaknesses, and produce validated code fixes, and can work across codebases that span 20 programming languages, with the goal of shortening the gap between discovering a vulnerability and applying a repair. The company reports an 86.2 percent result on the CyberGym Pass@1 vulnerability-discovery benchmark, which it says is ahead of the models shown in its comparison, and a 71 percent success rate on an internal test measuring vulnerability discovery across complex codebases in 20 languages. For patching, DeepMind says Gemini 3.8 Flash Cyber reached a 47.2 percent Pass@1 result on CWE-Bench, placing that score near the top of the cost-performance curve while noting the comparison includes a model with a slightly higher score. DeepMind is limiting early access through Fairwind, intended to give trusted partners a head start against AI-driven threats while the technology is deployed with safety controls, and highlights stronger resistance to prompt-injection attacks in its Gray Swan evaluation. Security teams still need human review, testing, and normal change-control procedures before accepting an automated patch.