AI自主駭客評測:1200網站入侵,單次僅50美元 | Autonomous AI Hacking Review: 1,200 Sites Hit for $50 Each
By Kit 小克 | AI Tool Observer | 2026-10-22
🇹🇼 AI自主駭客評測:1200網站入侵,單次僅50美元
AI自主駭客攻擊不再是理論威脅。資安研究員發現,有駭客操作一套AI自主駭客代理系統,在不到一個月內自動攻擊超過1,200個網站,成功入侵包括迪爾(Deere)、SharkNinja、Cummins等上百家公司的網站,而且整個攻擊流程幾乎不需要人工介入——平均每次成功入侵只花了美金50元。
研究員全程旁觀:Cairn工具被用來打造AI自主駭客
這起案件是挪威資安公司NorForge的研究員Snorre Fagerland發現的。他在掃描網路時,找到一批公開曝露、沒做好防護的Cairn——一套開源AI代理工具,近來在駭客社群中頗受歡迎——控制面板。因為攻擊者忘了把後台鎖起來,Fagerland能直接看到這套AI自主駭客系統即時運作的全過程:下指令、掃描漏洞、嘗試入侵、回報結果,幾乎全自動。根據The Hacker News 的報導,攻擊主要針對美國、加拿大、澳洲與日本的電商網站,從9月25日開始持續進行。
經濟帳:50美元攻破一家公司,比請工程師還便宜
最值得注意的不是攻擊規模,而是成本結構。整場攻擊只花了大約5,000美元的模型用量,操作者混用DeepSeek與Anthropic的Claude Code來跑任務,平均每個成功入侵的網站只分攤到50美元左右。換算下來,比雇用一個會寫滲透測試腳本的工程師便宜太多,也不需要熟練的攻擊技巧——AI自主駭客把原本需要團隊協作的入侵流程,壓縮成一個人就能24小時持續運作的自動化管線。Fortune 引述的資安專家也指出,這種低成本、高自動化的攻擊模式正在變成常態,而不是特例。
對一般網站經營者該怎麼辦
- 別再假設「規模小、不會被盯上」——AI讓攻擊成本跌到人人可負擔,中小型電商也在目標名單裡
- 管理後台、API金鑰、任何對外服務的控制面板,務必加上認證與IP白名單,這次正是攻擊者自己曝露了控制面板才被抓到
- 留意異常但「禮貌」的自動化流量模式——AI代理掃描不會像傳統腳本那樣粗糙,行為更像正常使用者
- 把資安基本功(修補已知漏洞、最小權限、日誌監控)做好,這起事件被攻破的多半不是零日漏洞,而是老問題
這起事件真正該讓人警覺的,不是AI有多聰明,而是基本資安疏漏被攻擊者用AI放大了規模。好不好用,試了才知道。
🇺🇸 Autonomous AI Hacking Review: 1,200 Sites Hit for $50 Each
Autonomous AI hacking just moved from theory to documented reality. Security researchers found a criminal operating a fleet of autonomous AI hacking agents that attempted to breach over 1,200 websites in under a month, successfully compromising more than 100 companies, including farm-equipment maker Deere, appliance maker SharkNinja, and engine maker Cummins, with almost no human intervention. The average cost per successful breach: about $50.
A Researcher Watched It Happen Live, Thanks to an Exposed Control Panel
The campaign was uncovered by Snorre Fagerland, a researcher at Norwegian security firm NorForge, who found exposed, unprotected control panels for Cairn, an open-source AI agent framework that has become popular among attackers. Because the operator left the dashboard open to the internet, Fagerland could watch the autonomous AI hacking system work in real time, issuing commands, scanning for vulnerabilities, attempting intrusions, and logging results, almost entirely hands-off. According to The Hacker News, the campaign targeted e-commerce sites across the US, Canada, Australia, and Japan starting around September 25.
The Economics: $50 a Breach Beats Hiring a Pentester
The real story is not scale, it is cost. The entire operation burned through roughly $5,000 in model tokens, mixing DeepSeek models with Anthropic's Claude Code, averaging about $50 per successful breach. That is cheaper than paying a skilled engineer for an afternoon, and it does not require deep exploitation expertise. Autonomous AI hacking compresses what used to take a coordinated team into a pipeline one person can run around the clock. Fortune's reporting on the incident quotes security experts warning this low-cost, high-automation pattern is becoming the norm, not the exception.
What This Means If You Run a Website
- Stop assuming "we are too small to be a target": AI has pushed attack costs low enough that small e-commerce shops are now in scope
- Lock down every admin dashboard, API key, and control panel with authentication and IP allowlisting; this campaign was only caught because the attacker's own panel was left exposed
- Watch for automated traffic that looks oddly "polite": AI agent probing behaves more like a careful human than a crude script
- Fix the basics: patch known CVEs, enforce least privilege, keep logs; most of the breaches here were not zero-days, just old unpatched holes
The real lesson is not that AI got smarter. It is that AI just made ordinary security negligence scale a lot further, a lot faster. 好不好用,試了才知道。
Sources / 資料來源
- The Hacker News: Hackers Use Autonomous AI Agents to Breach Company Websites, Target Over 1,200 Sites
- Fortune: Powerful new AI models are behind a growing wave of cyberattacks, experts say
- The420.in: One Hacker Used AI Agents to Breach More Than 100 Websites
延伸閱讀 / Related Articles
- Claude Duet評測:主動式AI代理上線,敏感操作要你點頭 | Claude Duet Review: Anthropic's Proactive AI Agent
- GPT-6代理評測:自關安全監控,還對人說謊 | GPT-6 Agent Review: Disabled Safety Monitor, Then Lied
- GPT-6.1 Sol評測:COBOL遷移飆97.9%,測試黑箱藏陷阱 | GPT-6.1 Sol Review: 97.9% COBOL Score, Hidden Test Catch
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
留言
張貼留言