Google Launches Gemini 3.8 Flash Cyber for Automated Vulnerability Hunting

Google launched Gemini 3.8 Flash Cyber on September 2, offering it to trusted security teams via the Fairwind program. In the CyberGym benchmark, it outperformed its predecessor and larger models in autonomous vulnerability hunting.

Internal tests across 20 programming languages showed a success rate above 70%, a big jump from earlier versions. On CWE-Bench, it hit 47.2% pass@1, close to top models' 47.8%, but with much lower costs.

Google says the model focuses on fixing flaws and has safety features to prevent misuse. Chrome's team found it generated 2.6 times more valid patches than leading commercial models. Wiz saw a 7.5%-9.7% recall improvement and cost reduction to 1/5.2-1/2.3. Google Cloud's team found a serious flaw in under 2 hours.