Introducing Gemini 3.5 Flash Cyber

Google has introduced Gemini 3.5 Flash Cyber, a new lightweight cybersecurity model built on top of 3.5 Flash and fine-tuned to quickly and efficiently find, validate, and patch software vulnerabilities. Because AI agents are becoming capable of finding vulnerabilities faster than defenders can fix them, the new model provides a scalable and cost-efficient alternative to massive, costly cybersecurity models. Due to the dual-use nature of this technology, Google is initially deploying the model through a limited-access pilot program where it will be exclusively available to governments and trusted partners via the CodeMender code security agent, with plans to expand over time to mitigate misuse.

Finding deep-seated software flaws requires exploring a massive execution search space, which can create a bottleneck for expensive, single-call large language models. CodeMender solves this by invoking 3.5 Flash Cyber multiple times, allowing sub-agents to analyze vast amounts of code paths and generate a single high-quality report. This speed and affordability make the model suitable for frequent scans, time-sensitive launches, and commit scanning pipelines at scale.

In benchmark testing on the CyberGym evaluation, CodeMender configured to call 3.5 Flash Cyber up to five times achieved performance competitive with significantly larger models. Additionally, stress tests by Google’s Big Sleep team on complex codebases like Chrome and Safari showed that 3.5 Flash Cyber significantly outperformed mainline 3.5 and 3.6 Flash models. In Google Chrome production pipeline evaluations and V8 JavaScript Engine tests, 3.5 Flash Cyber discovered more unique vulnerabilities than mainline 3.5 Flash and Claude Opus 4.6.

The model is already being used inside Google to secure internal codebases including Chrome, Android, Cloud, Ads, and YouTube. For instance, Google's Cloud Vulnerability Research team used the model to uncover remote code execution and memory-corruption vulnerabilities in public APIs and sensitive production services in just two hours, while also generating reliable exploits that bypassed standard mitigations.