Introducing Gemini 3.5 Flash Cyber

What it claims

Gemini 3.5 Flash Cyber is a fine-tune of Gemini 3.5 Flash for finding, validating and patching vulnerabilities. Google says that fanning out many cheap calls beats one expensive large-model call when an agent must scan a big codebase. On V8, with a fixed number of invocations, it found 55 unique confirmed issues against 47 for mainline 3.5 Flash and 36 for Claude Opus 4.6. Ten of those 55 were caught by neither of the other two models. Google also says it keeps finding new code paths and vulnerabilities as the invocation count grows, where weaker models "get stuck in a loop" re-finding the same bug.

How it works

• Orchestration. CodeMender calls the model several times as sub-agents covering different code paths. Their findings are merged into a single report. In the CyberGym run this was capped at five calls per final report.
• Training data. The page names OSV.dev (over 700,000 open-source vulnerabilities) and more than 10 years of OSS-Fuzz results as the sources of "real" vulnerability examples, used instead of synthetic ones.
• Training targets. The model is trained to operate industry-standard security tools and to work through multi-hour analyses of very large codebases such as Chromium.
• Not disclosed. The page gives no model size, training compute, or fine-tuning method.

Evidence

The text gives no figures for any of these three; numbers appear only in charts. The V8 counts above are the only hard numbers on the page.

Anecdote. Google's Cloud Vulnerability Research team found remote code execution bugs in public APIs and a memory-corruption bug in a production service within 2 hours. The model then produced an exploit described as "100% reliable" that bypassed ASLR and W^X.

Deployment. Already in internal use on Chrome, Android, Cloud, Ads and YouTube. Testers at Wiz and Cloud CISO Security Engineering report gains over mainline 3.5 Flash.

How to use it

You likely cannot use it yet:

• Access. It will be available only through CodeMender, in a limited pilot for governments and trusted partners. The page says "soon" and gives no date.
• Interfaces. No API, model ID, weights or pricing are given.
• Alternative. CodeMender's "foundational capabilities" are being offered separately on generally available Gemini models via the Gemini Enterprise Agent Platform. Note that this is not the Cyber model.

Limits

• CyberGym comparisons mix Google's own runs with competitors' self-reported scores.
• Chrome pipeline comparisons leave out Anthropic models newer than Opus 4.6, because those refused the tasks. The comparison set is therefore shaped by competitors' guardrails, not by capability.
• Big Sleep results were obtained with safety guardrails off. Production behaviour with guardrails on is not reported.
• V8 comparison: the "fixed number of invocations" is not stated, so cost-matched efficiency cannot be checked.
• Pilot access is restricted because of explicit dual-use concerns. The page itself shows the model writing a working mitigation-bypassing exploit.