Gemini 3.5 Flash Cyber評測：Google限量開放抓漏AI | Gemini 3.5 Flash Cyber Review: Google's Gated Bug Hunter

7月 28, 2026

By Kit 小克 | AI Tool Observer | 2026-07-28

🇹🇼 Gemini 3.5 Flash Cyber評測：Google限量開放抓漏AI

什麼是 Gemini 3.5 Flash Cyber？

Google DeepMind 在 2026 年 7 月 22 日發布 Gemini 3.5 Flash Cyber，這是專門用來挖漏洞的輕量版模型，架構奠基於 Gemini 3.5 Flash，但目前只開放給政府與「受信任夥伴」透過 CodeMender 平台試用，一般開發者暫時碰不到。這款模型的重點不是聊天或寫程式，而是專精在找出、驗證、甚至協助修補軟體漏洞。

小模型打贏大模型的關鍵數字

最值得注意的是效能：Gemini 3.5 Flash Cyber 在 CyberGym 基準測試拿下 83.2% 的成績，跟 OpenAI、Anthropic 那些體積大上好幾倍、成本高得多的前沿級資安代理人不相上下（這些模型普遍落在 83%~85.6% 區間）。更具體的實測來自 V8 JavaScript 引擎：Gemini 3.5 Flash Cyber 抓出 555 個確認漏洞，同條件下無專精版的 Gemini 3.5 Flash 只找到 474 個，Claude Opus 4.6 更只有 363 個。就算把呼叫次數限制在同一水準，Gemini 3.5 Flash Cyber 仍以 55 比 47（Gemini 3.5 Flash）、36（Opus 4.6）勝出，其中還有 10 個漏洞是只有它抓得到的。

CodeMender 怎麼用它

Gemini 3.5 Flash Cyber 不是單獨跑一次就結案，而是被 CodeMender 重複呼叫最多五次，讓多個子代理人（sub-agent）各自分析不同的程式碼路徑，最後彙整成一份完整報告。這種「小模型、多次呼叫、彙整結果」的做法，某種程度上比單純堆大模型參數更划算，也呼應了業界最近討論的「用便宜模型堆疊取代單一昂貴模型」趨勢。

為什麼刻意限量開放？

Google 很直白地說了原因：一款擅長找漏洞的 AI，同時也是一款擅長替攻擊者找漏洞的 AI。這是典型的雙重用途（dual-use）困境，所以現階段 Gemini 3.5 Flash Cyber 只透過 CodeMender 提供給政府機關與受信任夥伴的封閉試點，沒有對外 API、也沒有申請管道。這跟先前 Anthropic 把 Mythos 提供給 CISA 稽核政府軟體的邏輯很類似——資安圈似乎正在形成一種默契：威力強大的攻防型 AI 模型，先鎖在政府與受信任圈子裡試跑，而不是直接開放給所有人。

對一般開發者跟資安團隊來說，現階段能做的就是持續觀察：如果 Gemini 3.5 Flash Cyber 證明能大量降低漏洞修補的時間成本，遲早會有下放版本或商業化管道出現，屆時才是真正檢驗它是否好用的時候。

🇺🇸 Gemini 3.5 Flash Cyber Review: Google's Gated Bug Hunter

What Is Gemini 3.5 Flash Cyber?

On July 22, 2026, Google DeepMind released Gemini 3.5 Flash Cyber, a specialized, lightweight variant of Gemini 3.5 Flash built specifically to hunt, validate, and help patch software vulnerabilities. It isn't a general chat or coding model — it's a narrow tool for one job: finding bugs before attackers do. And right now, you can't use it directly. Access runs exclusively through CodeMender, limited to governments and "trusted partners" in a closed pilot program.

A Small Model Beating Bigger Ones

The headline number is efficiency: Gemini 3.5 Flash Cyber scores 83.2% on the CyberGym benchmark, putting it in the same range as far larger, far more expensive frontier security agents from OpenAI and Anthropic, which cluster between 83% and 85.6%. The more concrete test came from the V8 JavaScript engine: Gemini 3.5 Flash Cyber found 555 confirmed unique issues, versus 474 for the base Gemini 3.5 Flash and just 363 for Claude Opus 4.6. Even under a fixed-invocation budget, it still won — 55 confirmed issues versus 47 (Gemini 3.5 Flash) and 36 (Opus 4.6), including 10 bugs no other model caught at all.

How CodeMender Uses It

The trick isn't a single pass. CodeMender calls Gemini 3.5 Flash Cyber up to five times per assessment, spinning up multiple sub-agents that each analyze different code paths before merging everything into one combined report. It's a "small model, called repeatedly, aggregated" strategy rather than brute-forcing with a bigger model — echoing a broader industry shift toward cheaper models stacked in agentic loops instead of single expensive calls.

Why the Gate?

Google is upfront about the reasoning: a model good at finding vulnerabilities is, by definition, also good at finding them for attackers. That dual-use tension is why Gemini 3.5 Flash Cyber has no public API and no signup form — it's confined to a closed CodeMender pilot for governments and vetted partners. It echoes Anthropic handing Mythos to CISA for government code audits: the security world seems to be settling into a pattern where powerful offense-capable AI ships to trusted, government-adjacent circles first, not the open market.

For most developers and security teams, there's nothing to test yet — just something to watch. If Gemini 3.5 Flash Cyber proves it meaningfully cuts patch time during the pilot, a broader or commercial rollout is the logical next step, and that's when it's actually worth trying.

Sources / 資料來源

延伸閱讀 / Related Articles

AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends

搜尋此網誌

AI小貼士