Google has introduced Gemini 3.5 Flash Cyber, a lightweight cybersecurity model built on the 3.5 Flash architecture and fine-tuned to find, validate, and help patch software vulnerabilities. The company positions the model as a fast, cost-efficient option for defenders who must scan large codebases and explore many code paths.
Why a lightweight model matters
Automated vulnerability discovery must search a vast execution space. Using a single expensive call to a large language model can become a bottleneck. By contrast, a lightweight model like 3.5 Flash Cyber can be invoked multiple times across many code paths, enabling agents to analyze far more execution paths and then aggregate sub-agent results into a single, high-quality report. Google says the model’s speed and affordability make it suitable for frequent scans, time-sensitive launch processes, and commit-scanning pipelines at scale.
Access and deployment approach
Because this technology has dual-use risks, Google is rolling out 3.5 Flash Cyber deliberately. The model will be available initially through a limited-access pilot program via CodeMender, exclusively to governments and trusted partners, and will expand over time. Google says this approach aims to give frontline defenders an advantage in finding and fixing critical vulnerabilities while mitigating broader misuse. Separately, CodeMender’s core capabilities will be made available more broadly to customers through generally available Gemini models on the Gemini Enterprise Agent Platform.
Benchmark and evaluation results
Google tested 3.5 Flash Cyber on several benchmarks. On CyberGym, which evaluates AI agents against hundreds of real-world software vulnerabilities, Google configured CodeMender to call 3.5 Flash Cyber up to five times per final report; with that setup, the overall agent achieved competitive performance versus significantly larger models (competitor results are provider self-reported).
Google’s Big Sleep team ran independent evaluations focused on hard-to-find vulnerabilities in complex codebases such as Chrome and Safari. In that evaluation, 3.5 Flash Cyber significantly outperformed mainline 3.5 Flash and 3.6 Flash on the Big Sleep pass@1 metric.
The model was also evaluated in Google Chrome’s production commit scanning pipeline. The vulnerabilities found were not publicly disclosed to prevent contamination of the benchmark for Gemini and competitor models. Results reportedly showed a significant uplift for 3.5 Flash Cyber compared to 3.5 Flash. Google notes that some competitor models released after Opus 4.6 refused to perform the tasks due to built-in safety guardrails and therefore are not shown.
In an additional test on the V8 JavaScript Engine with a fixed number of invocations, 3.5 Flash Cyber found 55 unique confirmed issues, compared with 47 found by mainline 3.5 Flash and 36 by Opus 4.6. Ten of the issues found by 3.5 Flash Cyber were not discovered by the other two models. Google emphasizes that basic cybersecurity models can loop on the same findings and miss critical issues; a stronger model should discover a higher number of unique problems. As the number of invocations scales up, 3.5 Flash Cyber continued to discover new code paths and vulnerabilities.
Real-world use at Google
Beyond benchmarks, Google reports that 3.5 Flash Cyber within CodeMender is already used to find and fix vulnerabilities in internal codebases including Chrome, Android, Google Cloud, Ads, and YouTube. Google provides an example from its Cloud Vulnerability Research team: in about two hours the model uncovered remote code execution vulnerabilities in public APIs and a memory-corruption bug in a sensitive production service, and it produced a reliable remote-code-execution exploit that bypassed mitigations such as Address Space Layout Randomization (ASLR) and Write XOR Execute (W^X). Early feedback from testers at Wiz and the Cloud CISO Security Engineering team reportedly confirms a notable capability improvement of 3.5 Flash Cyber over the mainline 3.5 Flash model.
Training data and defender enablement
Google cites resources that support model training on real-world examples: OSV.dev, the vulnerability database run by Google that covers over 700,000 open-source vulnerabilities, and more than ten years of OSS-Fuzz results. These datasets are used to teach models to use industry-standard tools, analyze millions of lines of code in large projects like Chromium, and perform complex, hours-long analysis tasks independently.
Conclusion
Gemini 3.5 Flash Cyber is a purpose-built, lightweight model intended to give defenders a scalable, cost-effective tool for vulnerability discovery and validation. Google plans a cautious, limited rollout through CodeMender for governments and trusted partners while broader capabilities from CodeMender will be exposed via generally available Gemini models.
