Google DeepMind has introduced Gemini 3.5 Flash Cyber, a specialized version of its lightweight Flash model designed for cybersecurity work. The model is fine-tuned to help find, validate, and patch software vulnerabilities faster and more cheaply than larger general-purpose models used for the same tasks.
The company says the model is meant for workflows where an agent must inspect large codebases and analyze many possible execution paths. Because Flash is smaller and less costly to run, CodeMender can call Gemini 3.5 Flash Cyber many times through sub-agents and combine their work into a final vulnerability report.
Access will be deliberately narrow at first. Google says the model will be available soon through a limited pilot for governments and trusted partners via CodeMender, expanding over time. The company frames that restriction as a way to help defenders while reducing misuse risk from a tool that could also aid attackers.
The release shows how AI labs are moving from broad coding assistants toward task-specific security models. Its usefulness will depend not only on benchmark results, but on whether controlled access and deployment practices keep the tool on the defensive side.