Google DeepMind has announced Gemini 3.5 Flash Cyber, a specialized cybersecurity model built on the Gemini 3.5 Flash architecture. Designed to find, validate, and patch software vulnerabilities efficiently, the model aims to give defenders a scalable and cost-effective tool against rising AI-powered threats.
Key Features and Deployment
Gemini 3.5 Flash Cyber is fine-tuned for code security tasks, outperforming mainline Flash models in vulnerability detection. It is integrated into CodeMender, Google's automated code security agent, which can invoke the model multiple times to analyze vast codebases. The model is initially available via a limited-access pilot program to governments and trusted partners, with broader release planned over time.
Benchmark Performance
On the CyberGym benchmark, CodeMender using 3.5 Flash Cyber achieved competitive results against larger models. In Google's internal Big Sleep evaluation, it significantly surpassed Gemini 3.5 Flash and 3.6 Flash in finding critical vulnerabilities in complex codebases like Chrome and Safari. On Chrome's production commit scanning pipeline, it showed a significant uplift over 3.5 Flash and discovered more unique vulnerabilities than Claude Opus 4.6 in the V8 JavaScript Engine (55 vs. 47 and 36).
Real-World Impact
Google's internal teams have used 3.5 Flash Cyber to find remote code execution and memory-corruption vulnerabilities in production services within hours. The model is already deployed across Chrome, Android, Cloud, Ads, and YouTube codebases. Google emphasizes that training on real-world vulnerabilities from OSV.dev and OSS-Fuzz helps the model learn from actual security workflows.
Safety and Access
Given the dual-use nature, Google is restricting initial access to trusted entities to mitigate misuse while empowering defenders. The Gemini Enterprise Agent Platform also brings CodeMender capabilities to general customers.