GPT-5.6-Cyber評測：OpenAI攻防級AI找出Chrome零時差漏洞 | GPT-5.6-Cyber Review: OpenAI Offense-Grade AI Finds Chrome Zero-Days

By Kit 小克 | AI Tool Observer | 2026-08-17

🇹🇼 GPT-5.6-Cyber評測：OpenAI攻防級AI找出Chrome零時差漏洞

GPT-5.6-Cyber 是 OpenAI 於 2026 年 8 月推出的專用資安模型，也是第一個在 OpenAI 自家「Preparedness Framework」風險評估中被歸類為「High」網路安全能力等級的模型。它已經在 Chrome 的 V8 引擎裡挖出真實可用的零時差漏洞（CVE-2026-15903），而且不是玩具案例——是能串連成完整沙盒逃逸的那種。

什麼是 Daybreak Red？門檻比一般 ChatGPT 高很多

GPT-5.6-Cyber 不是隨便誰都能用。OpenAI 把它放進代號「Daybreak」的分級存取計畫，分成兩層：

Daybreak Blue：一般核准的防守方可以用 GPT-5.6 Sol 等通用旗艦模型做授權防禦工作
Daybreak Red：更高門檻，專門開放給做漏洞研究、滲透測試、exploit 驗證的資安團隊，GPT-5.6-Cyber 就在這一層

關鍵數字：在內部測試中，面對進階資安請求，一般版 GPT-5.6 Sol（Daybreak Blue）只有 2% 會直接執行，其餘幾乎都被安全機制擋下來；換成 GPT-5.6-Cyber，完成率飆到 95%。這代表 OpenAI 刻意把這個版本的拒答門檻拉低，換取更高的實戰可用性。

找到什麼：V8 引擎的記憶體越界漏洞

GPT-5.6-Cyber 挖出的 CVE-2026-15903 出在 V8 的 TurboFan 最佳化編譯器：在特定條件下，將數值轉成整數時漏掉了一個安全檢查，讓攻擊者能對陣列邊界外的記憶體做讀寫。更值得注意的是，它還找到了第二個漏洞，能把這個「堆積讀寫原語」串接成完整的沙盒逃逸——OpenAI 至今沒公開第二個 CVE 編號。Google 已經在 Chrome 150.0.7871.128 修補了第一個漏洞。

這不是 OpenAI 第一次碰上這類發現：Google 自家的 DeepMind「Big Sleep」計畫也用 AI agent 在 V8 裡挖過漏洞，說明找漏洞正在變成 AI agent 的固定戰場，不再是單一實驗室的專利。

這對開發者跟企業意味著什麼

防守窗口正在縮短：OpenAI 自己把這次發表定調為「Cyber Defense Window Narrows」——意思是漏洞從被發現到被武器化的時間差正在被 AI 壓縮，企業修補速度必須跟上
紅隊測試門檻降低：如果你的團隊做滲透測試或漏洞賞金，Daybreak Red 值得申請；但要注意這仍需要 OpenAI 審核資格
雙面刃現實：同樣的能力落到未經授權的手上就是攻擊工具，OpenAI 把存取權限卡在審核制，但業界對這類模型的外流風險仍有疑慮

誠實講：值得關注但別神化

GPT-5.6-Cyber 目前仍低於 OpenAI 定義的「Critical」門檻，也就是還沒到「能完全自動化端到端攻擊已強化目標」的程度。它更像是資深資安研究員的加速器，而不是萬能駭客機器人。如果你是防守方，重點不是恐慌，而是確認自己的修補流程夠快、依賴的第三方套件有沒有持續掃描。

好不好用，試了才知道。

🇺🇸 GPT-5.6-Cyber Review: OpenAI Offense-Grade AI Finds Chrome Zero-Days

GPT-5.6-Cyber is OpenAI's dedicated cybersecurity model, launched in August 2026, and the first model OpenAI itself classified as "High" cyber capability under its Preparedness Framework. It has already found a real, working zero-day in Chrome's V8 engine (CVE-2026-15903) — and not a toy bug either, one that chains into a full sandbox escape.

Daybreak Red: A Much Higher Bar Than Regular ChatGPT

GPT-5.6-Cyber isn't open to just anyone. OpenAI gates it behind a program called Daybreak, split into two tiers:

Daybreak Blue: approved defenders get access to general flagship models like GPT-5.6 Sol for authorized defensive work
Daybreak Red: a higher bar reserved for teams doing vulnerability research, exploit validation, and pentesting — this is where GPT-5.6-Cyber lives

The telling number: in internal evals, standard GPT-5.6 Sol under Daybreak Blue completed only 2% of advanced cybersecurity requests, refusing the rest. GPT-5.6-Cyber completes 95%. OpenAI deliberately lowered the refusal threshold on this variant in exchange for real operational usefulness.

What It Found: A Memory Out-of-Bounds Bug in V8

CVE-2026-15903, the bug GPT-5.6-Cyber uncovered, sits in V8's TurboFan optimizing compiler: under specific conditions, a safety check was skipped when converting values to integers, letting an attacker read or write memory outside array bounds. More notably, the model also found a second, still-unnamed vulnerability that turns this heap read/write primitive into a full sandbox escape. Google has already patched the first bug in Chrome 150.0.7871.128.

This isn't an isolated OpenAI moment — Google's own DeepMind "Big Sleep" project has used AI agents to find bugs in V8 too, suggesting automated vulnerability discovery is becoming a standard battlefield for AI agents across labs, not one company's edge.

What This Means for Developers and Enterprises

The defense window is shrinking: OpenAI framed the launch itself as the "cyber defense window narrowing" — the gap between a bug being found and weaponized is compressing, and patch cycles need to keep up
Red-teaming just got easier to access: if your team does pentesting or bug bounty work, Daybreak Red is worth applying for — though OpenAI still vets eligibility
The dual-use reality: the same capability in unauthorized hands is an attack tool; OpenAI gates access behind review, but the industry still worries about leakage risk for models like this

Honest Take: Worth Watching, Not Worth the Hype

GPT-5.6-Cyber still sits below OpenAI's "Critical" threshold — it can't yet fully automate end-to-end attacks against hardened targets. Think of it as an accelerant for a skilled security researcher, not an autonomous hacking machine. If you're on defense, the move isn't panic — it's checking whether your patch cadence and third-party dependency scanning can keep pace.

好不好用，試了才知道。

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

Claude Code 實測：AI 幫你寫程式到底行不行？ | Claude Code Review: Can AI Really Code for You?

3月 26, 2026

閱讀完整內容

搜尋此網誌

AI小貼士