Google DeepMind Launches Gemini 3.5 Flash Cyber to Hunt Code Vulnerabilities

Google DeepMind announced Gemini 3.5 Flash Cyber on July 21, 2026, a variant of the 3.5 Flash model fine-tuned specifically for finding, verifying, and patching software vulnerabilities. The model is currently available only to governments and trusted partners as part of a limited pilot program run through the CodeMender agent.

What Flash Cyber Is

Flash Cyber isn't a new standalone model built from scratch, but a version of Gemini 3.5 Flash fine-tuned for a specific task: analyzing code for vulnerabilities. Instead of a single costly query to a large model, CodeMender calls Flash Cyber repeatedly, scanning successive code execution paths, and finally compiles the results into one report.

This approach makes economic sense: a cheap, fast model run repeatedly can scan far more code than a single, expensive query to a large general-purpose model. Google says that thanks to this architecture, CodeMender achieves results competitive with much larger models on the CyberGym benchmark.

Test Results

The most concrete data comes from a test on the V8 JavaScript engine, one of the most complex components of web browsers. With the same number of calls, Flash Cyber found 55 unique, confirmed issues, while standard Gemini 3.5 Flash detected 47 and Anthropic's Claude Opus 4.6 found 36. Notably, 10 of the vulnerabilities found were missed by both of the other two models.

Google also reports results from Big Sleep tests on Chrome and Safari code, where Flash Cyber clearly outperformed both the mainline 3.5 Flash model and the newer 3.6 Flash. The company's Cloud Vulnerability Research team used the model to analyze production systems and within two hours found remote code execution vulnerabilities and memory corruption bugs, generating working exploits that bypass standard ASLR and W^X protections.

This will give defenders a critical time advantage in finding and fixing critical vulnerabilities before they can be exploited, while limiting the risk of broader misuse - Raluca Ada Popa and Four Flynn, Google DeepMind

Who Gets Access

Google is deliberately limiting access to Flash Cyber. The model is going exclusively to governments and trusted partners as part of a pilot run by CodeMender, Google's agent for automatically detecting and patching vulnerabilities in code. The company says it plans to gradually expand access, but for now, as the announcement put it, priority goes to frontline defenders, meaning the security teams of large organizations and public institutions.

The reason for this caution is obvious: the same model that can find a vulnerability in code can, in the wrong hands, be used to exploit it. Google openly acknowledges that this is fueling a race in which AI agents are becoming increasingly skilled at finding vulnerabilities faster than defenders can patch them, which in itself justifies a controlled-access model rather than public release of the tool.

CodeMender and the Bigger Picture

CodeMender is a Google system unveiled in October 2025 that automates the detection and repair of security vulnerabilities across large codebases, including Chrome, Android, Google Cloud, Ads, and YouTube. Flash Cyber is meant to power this infrastructure with a cheaper, faster analytical engine, allowing more code to be scanned without a proportional increase in computing costs.

Google is also making CodeMender's basic capabilities available to a wider range of companies through the Gemini Enterprise Agent platform, based on standard Gemini models, without access to the specialized Flash Cyber. This distinction shows that the company considers the most effective variant of the model too sensitive at this stage for general distribution.

What It Means for the Security Industry

For security teams in Poland and around the world, the announcement confirms the direction the entire industry is heading: vulnerability analysis increasingly relies on cheap, specialized models run at scale rather than single queries to the most expensive flagship systems. The comparison with Anthropic's Claude Opus 4.6 also shows that in narrow, well-defined tasks, a smaller, fine-tuned model can outperform a much more expensive general-purpose model.

Limited access means, however, that Polish companies and institutions won't be able to use Flash Cyber directly for now, unless they become part of the trusted partner circle in the pilot program. The general capabilities of CodeMender remain available through the Gemini Enterprise Agent platform, based on standard Gemini models, without the vulnerability-detection edge the Cyber variant provides.