Google has announced Gemini 3.5 Flash Cyber, a lightweight AI model designed for cybersecurity tasks such as vulnerability detection and patching. This model aims to enhance security operations with efficient, automated threat response.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
Google DeepMind releases Gemini 3.5 Flash Cyber, a lightweight model fine-tuned for cybersecurity. Through multiple invocations, it achieves vulnerability discovery capabilities comparable to larger models and has already made a real impact on Google's internal codebases.
Discovering deep-seated flaws requires exploring a vast execution search space. Relying on a single expensive large model call can create a bottleneck. 3.5 Flash Cyber is particularly suited for scenarios that require scanning large codebases and analyzing numerous code paths.
CodeMender invokes 3.5 Flash Cyber multiple times, enabling the agent to analyze more code paths to discover and verify vulnerabilities, then generate a single high-quality report. Its speed and cost-effectiveness make it easy to integrate into frequent scanning, time-sensitive release processes, or commit scanning pipelines.
On the CyberGym benchmark, with up to five CodeMender invocations of 3.5 Flash Cyber, the overall agent achieved performance competitive with significantly larger models. Note: Competitor results are from provider self-reports.
In evaluations independently built by the Big Sleep team, focusing on critical vulnerabilities in complex codebases like Chrome and Safari, 3.5 Flash Cyber significantly outperformed the baseline 3.5 Flash and 3.6 Flash.
In evaluations on the Google Chrome production commit scanning pipeline, with vulnerabilities undisclosed to avoid contamination, results showed significant improvements of 3.5 Flash Cyber over 3.5 Flash. Newer competitor models (e.g., after Opus 4.6) are not shown because they refused to execute tasks due to built-in safety guardrails.
In fixed-invocation tests on the V8 JavaScript engine, 3.5 Flash Cyber found 55 unique confirmed issues, while baseline 3.5 Flash found 47 and Claude Opus 4.6 found 36. Among these, 10 issues were not found by the other two models.
With increased invocations, 3.5 Flash Cyber consistently discovers new code paths and vulnerabilities, avoiding the base model's tendency to cycle through rediscovering the same issues.
3.5 Flash Cyber has already discovered and fixed vulnerabilities in Google's internal codebases (including Chrome, Android, Cloud, Ads, and YouTube) via CodeMender.
Google Cloud's vulnerability research team used the model to discover a remote code execution vulnerability in a public API and a memory corruption vulnerability in a sensitive production service within 2 hours, and generated 100% reliable remote code execution exploits that bypass standard mitigations like ASLR and W^X.
Early feedback from testers at Wiz and Cloud CISO Security Engineering confirms significant capability improvements of 3.5 Flash Cyber over baseline 3.5 Flash.
Given the dual-use nature of the technology, Google is taking a cautious deployment approach. As part of a limited-access pilot program, 3.5 Flash Cyber will first be exclusively available to governments and trusted partners via CodeMender, with gradual expansion.
This aims to give frontline defenders a head start in discovering and fixing critical vulnerabilities while mitigating broader abuse risks. Additionally, Google is making CodeMender's underlying capabilities directly available to customers through the Gemini Enterprise Agent Platform.
This article's information primarily comes from Google DeepMind's official press release, a first-party source. In the CyberGym benchmark, competitor results are from provider self-reports, which are less reliable. Other internal evaluations (e.g., Big Sleep, Chrome pipeline) were conducted by Google teams and may have bias. Real-world application cases come from internal Google teams and lack independent verification.
Gemini 3.5 Flash Cyber demonstrates the potential of lightweight models in cybersecurity: through multiple invocations, it achieves vulnerability discovery capabilities comparable to larger models while maintaining cost-effectiveness. It has already made a real impact inside Google, but deployment is limited to a pilot program, and performance data primarily comes from Google's own evaluations, requiring independent verification.
Primary report
Primary source