介绍 Gemini 3.5 Flash Cyber

2026年7月21日 模型

Raluca Ada Popa 与 Four Flynn

多年来，谷歌持续投资网络安全，率先开展自动化漏洞发现，以保障全球代码库的安全。诸如我们的代码安全智能体 CodeMender 等工具，能够自动发现并修复关键软件漏洞。然而，随着AI智能体发现漏洞的速度越来越快，甚至超过防御者修复的速度，应对这一全球性威胁需要一种高性能、低成本且可扩展的方案。

今天，我们通过推出 Gemini 3.5 Flash Cyber 来拓展长期以来的努力，更好地武装防御者。这是一款基于 3.5 Flash 构建的轻量级网络安全模型，经过微调后可快速高效地发现、验证并修补漏洞，在这些任务上的表现优于 Gemini 主线的 Flash 模型。

Flash 的性能与效率使其成为我们网络安全模型工作的理想基础。基于 Flash 构建的 3.5 Flash Cyber，为庞大且昂贵的网络安全模型提供了一种成本效益高且性能强大的替代方案。

鉴于该技术的双重用途性质，我们在部署 3.5 Flash Cyber 时采取了审慎的方式。作为有限访问试点计划的一部分，3.5 Flash Cyber 将很快通过 CodeMender 独家提供给政府和受信任的合作伙伴，并逐步扩大范围。这将使一线防御者在关键漏洞被利用之前，率先发现并修复它们，同时降低被滥用的风险。

此外，我们还通过 Gemini Enterprise Agent Platform，将 CodeMender 的基础能力直接以通用 Gemini 模型的形式提供给客户。

搜索空间问题：轻量级模型在代码安全中的优势

发现深层缺陷需要探索庞大的执行搜索空间。依赖单一、昂贵的大型模型调用可能形成瓶颈。3.5 Flash Cyber 特别适用于智能体需要扫描大型代码库并分析大量代码路径的漏洞发现场景。

CodeMender 会多次调用 3.5 Flash Cyber，使智能体能够分析更多的代码路径，从而发现并验证漏洞。随后，子智能体会生成一份高质量的综合报告。

凭借其速度和成本优势，3.5 Flash Cyber 可以轻松集成到频繁扫描、时间敏感的发布流程或大规模提交扫描管道中。

3.5 Flash Cyber 基准测试结果：大型网络安全模型的高效替代方案

我们在多种基准测试上对 3.5 Flash Cyber 进行了评估。特别是在 CyberGym 基准测试中，我们测试了 AI 智能体应对数百个真实世界软件漏洞的能力。通过利用 3.5 Flash Cyber 的低成本，配置 CodeMender 最多调用 5 次 3.5 Flash Cyber 来生成一份最终报告，整体智能体在 CyberGym* 上取得了与显著更大的模型相竞争的性能。

*竞争对手结果来源于提供商自行报告的数据

我们还在没有安全护栏的情况下，对模型在 CyberGym 之外的能力进行了压力测试。谷歌的 Big Sleep 团队独立构建了一项评估，专注于在 Chrome 和 Safari 等全球最复杂的代码库中发现关键且难以发现的漏洞。在此评估中，3.5 Flash Cyber 显著超越了主线的 3.5 Flash 和 3.6 Flash。

Big Sleep 评估成功率 (pass@1)

3.5 Flash Cyber 还在谷歌 Chrome 的生产提交扫描管道中接受了评估。这些漏洞并未公开披露，从而确保该基准测试对 Gemini 及竞争对手模型均无污染。

结果显示，与 3.5 Flash 相比，3.5 Flash Cyber 有显著提升。注意：Opus 4.6 之后更新的竞争对手模型版本因内置安全护栏而拒绝执行任务，因此未显示。

Chrome 生产提交扫描管道成功率 (pass@1)

此外，与主线 3.5 Flash 和 Claude Opus 4.6 相比，3.5 Flash Cyber 持续发现更多独特的漏洞。在对高度复杂的 V8 JavaScript 引擎进行固定次数调用的测试中，3.5 Flash Cyber 发现了 55 个已确认的独特问题，而主线 3.5 Flash 发现了 47 个，Opus 4.6 发现了 36 个，其中包括另外两个测试模型未捕获的 10 个问题。

基础的网络安全模型可能会陷入循环，重复发现相同的问题，同时遗漏关键漏洞。一个强大的模型能撒下更广的网，发现更多独特的问题。

随着我们增加调用次数，我们发现 3.5 Flash Cyber 会持续发现新的代码路径和漏洞。

谷歌内部的实际应用与规模化防御

基准测试只是故事的一部分。CodeMender 中的 3.5 Flash Cyber 已经在谷歌的内部代码库（包括 Chrome、Android、Cloud、Ads 和 YouTube）中发现并修复漏洞。

轻量级模型带来的发现速度已产生可衡量的影响。

例如，谷歌的云漏洞研究团队使用 3.5 Flash Cyber 以前所未有的速度主动保护我们的系统。在短短 2 小时内，该模型在公共 API 中发现了远程代码执行漏洞，并在一个敏感的生产服务中发现了内存损坏漏洞。随后，它生成了一个 100% 可靠的远程代码执行利用程序，绕过了地址空间布局随机化（ASLR）和写异或执行（W^X）等标准缓解技术。

来自 Wiz 和 Cloud CISO 安全工程测试人员的早期反馈证实，3.5 Flash Cyber 相比主线 3.5 Flash 模型在能力上有显著提升。

大规模赋能防御者

谷歌在软件安全领域的领导地位赋予了我们独特的优势。例如，OSV.dev（谷歌运营的漏洞数据库，涵盖超过 70 万个开源漏洞）以及超过 10 年的 OSS-Fuzz 结果，帮助我们识别最高质量的漏洞。

这使我们能够超越合成的网络安全示例，教会模型真实安全专业人员的工作方式。我们的模型学会操作行业标准工具，阅读 Chromium 等大型项目中数百万行代码，并独立处理需要数小时持续深度分析的复杂安全任务。

通过用 3.5 Flash Cyber 驱动 CodeMender，我们提供了一种高性能、可扩展且经济实惠的架构，旨在帮助更多防御者保障软件安全。

July 21, 2026 Models

Raluca Ada Popa and Four Flynn

Google has invested in cybersecurity for years, pioneering automated vulnerability discovery to secure the world’s codebases. Tools like CodeMender, our code security agent, can automatically find and fix critical software vulnerabilities. But as AI agents become more capable at finding vulnerabilities faster than defenders can fix them, addressing this global threat requires a highly capable, affordable, and scalable approach.

Today, we’re expanding our longtime efforts to better prepare defenders by introducing Gemini 3.5 Flash Cyber, our lightweight cybersecurity model built on top of 3.5 Flash and fine-tuned to find, validate, and patch vulnerabilities quickly and efficiently, making it more effective at these tasks than Gemini’s mainline Flash models.

Flash’s performance and efficiency makes it an ideal foundation for our cybersecurity model efforts. By building on top of Flash, 3.5 Flash Cyber offers a cost-efficient and highly capable alternative to large, costly cybersecurity models.

Given the dual-use nature of this technology, we have taken an intentional approach to how we deploy 3.5 Flash Cyber. As part of a limited-access pilot program, 3.5 Flash Cyber will be exclusively available to governments and trusted partners via CodeMender soon, expanding over time. This will give frontline defenders a head start in finding and fixing critical vulnerabilities before they can be exploited, while mitigating against broader misuse.

Separately, we're also bringing CodeMender's foundational capabilities directly to customers with generally available Gemini models through the Gemini Enterprise Agent Platform.

The search space problem: The advantage of lightweight models in code security

Finding deep-seated flaws requires exploring an immense execution search space. Relying on a single, expensive call to a massive language model can create a bottleneck. 3.5 Flash Cyber is particularly suitable for finding vulnerabilities where the agent has to scan a large codebase and analyze a large number of codepaths.

CodeMender invokes 3.5 Flash Cyber multiple times, so agents can analyze vastly more code paths to discover and validate vulnerabilities. The sub-agents then produce a single, high-quality report.

Thanks to its speed and affordability, 3.5 Flash Cyber can be easily integrated into frequent scans, time-sensitive launch processes or commit scanning pipelines at scale.

3.5 Flash Cyber benchmark results: an efficient alternative to larger cybersecurity models

We tested 3.5 Flash Cyber on a variety of benchmarks. In particular, we tested 3.5 Flash Cyber on the CyberGym benchmark, which evaluates AI agents against hundreds of real-world software vulnerabilities. Leveraging the low cost of 3.5 Flash Cyber by configuring CodeMender to call 3.5 Flash Cyber up to five times for a single, final report, the overall agent achieved competitive performance against significantly larger models on CyberGym*.

*Competitor results are sourced from provider self-reported scores

We also stress-tested the model’s capabilities beyond CyberGym without safety guardrails. Google’s Big Sleep team independently built an evaluation focused on finding critical and hard to find vulnerabilities in some of the world’s most complex codebases like Chrome and Safari. Here, 3.5 Flash Cyber significantly surpassed mainline 3.5 Flash and 3.6 Flash.

Success Rate on Big Sleep Evaluation (pass@1)

3.5 Flash Cyber was also evaluated on Google Chrome’s production commit scanning pipeline. The vulnerabilities were not publicly disclosed, which ensured this benchmark remained free of contamination for Gemini and competitor models.

The results showed a significant uplift from 3.5 Flash Cyber compared to 3.5 Flash. Note: More recent competitor model versions after Opus 4.6 refuse to fulfill the tasks due to built-in safety guardrails, and therefore are not shown.

Success Rate on Chrome Production Commit Scanning Pipeline (pass@1)

Moreover, 3.5 Flash Cyber consistently discovered more unique vulnerabilities compared with mainline 3.5 Flash and Claude Opus 4.6. When tested on the highly complex V8 JavaScript Engine across a fixed number of invocations, 3.5 Flash Cyber found 55 unique confirmed issues, compared to 47 found by mainline 3.5 Flash and 36 found by Opus 4.6, including 10 issues that the other two models tested did not catch.

Basic cybersecurity models can get stuck in a loop, finding the same issue repeatedly while missing critical vulnerabilities. A strong model casts a wider net, finding a higher number of unique issues.

As we scale the number of invocations, we find that 3.5 Flash Cyber continues to discover new code paths and vulnerabilities.

Real-world application and scaling defenses at Google

Benchmarks are only part of the story. 3.5 Flash Cyber in CodeMender is already finding and fixing vulnerabilities in Google’s internal codebases including Chrome, Android, Cloud, Ads, and YouTube.

The speed of discovery made possible by a lightweight model has delivered measurable impact.

For example, Google’s Cloud Vulnerability Research team used 3.5 Flash Cyber to proactively secure our systems in record time. In just 2 hours, the model uncovered remote code execution vulnerabilities in public APIs and found a memory-corruption vulnerability in a sensitive production service. It then generated a 100% reliable remote-code execution exploit that bypassed standard mitigation techniques like Address Space Layout Randomization (ASLR) and Write XOR Execute (W^X) .

Early feedback from Wiz and Cloud CISO Security Engineering testers confirms the significant capability improvement of 3.5 Flash Cyber over the mainline 3.5 Flash model.

Empowering defenders at scale

Google’s leadership in software security gives us a unique advantage. For example, OSV.dev, a vulnerability database run by Google spanning over 700,000 open-source vulnerabilities, and more than 10 years of OSS-Fuzz results, help us identify the most high quality vulnerabilities.

This allows us to move beyond synthetic cybersecurity examples and teach our models how real security professionals work. Our models learn to operate industry-standard tools, read through millions of lines of code in large-scale projects like Chromium, and independently tackle complex security tasks that require hours of continuous, deep analysis.

By powering CodeMender with 3.5 Flash Cyber, we’re providing a highly capable, scalable, and affordable architecture designed to help more defenders secure software.