AI token theft 指偷走或濫用可解鎖付費 AI 算力嘅 API key、session/OAuth token、automation credential 或免費試用額度。 常見兩大模式係大量假帳戶攞免費 credit,以及 LLMjacking:偷 AI API key 後用受害者帳戶跑模型、燒帳單 [1][4][5]。

Create a landscape editorial hero image for this Studio Global article: What is token theft in AI platforms, and why is it becoming a major fraud problem for AI startups?. Article summary: Token theft in AI platforms means stealing or abusing the “tokens” that grant access to AI compute—such as API keys, session tokens, free-trial credits, or prepaid usage credits. It is becoming a major fraud problem beca. Topic tags: general, general web, user generated. Reference image context from search candidates: Reference image 1: visual subject "An attacker sends a phishing email to users, steals credentials and tokens through a compromised MFA check, then replays a session token to gain access to a legitimate website by e" Reference image 2: visual subject "This technique involves adversaries stealing application access tokens, such as account API tokens, to gain unauthorized access to remote s
AI token theft 可以理解為偷 AI 平台「計費層」嘅錢。攻擊者未必偷到模型本身,真正值錢嘅係可以使用模型嘅存取權:例如 API key、OAuth token、session token、automation credential,甚至係註冊時送出嘅免費試用 credit [3][
6]。
換句話講,黑產偷嘅係「可以燒算力嘅通行證」。一旦得手,佢哋就可以用受害者或平台嘅帳戶跑模型、消耗昂貴 compute,最後由 AI 公司或合法帳戶持有人埋單 [1][
4][
5]。
講 AI 時,token 好容易令人諗起大語言模型處理文字時用嚟計量同收費嘅「文字片段」。但喺 token theft 呢個語境,token 通常指更闊嘅概念:任何可以打開付費用量嘅憑證或額度。
常見目標包括:
Studio Global AI
Use this topic as a starting point for a fresh source-backed answer, then compare citations before you share it.
AI token theft 指偷走或濫用可解鎖付費 AI 算力嘅 API key、session/OAuth token、automation credential 或免費試用額度。
AI token theft 指偷走或濫用可解鎖付費 AI 算力嘅 API key、session/OAuth token、automation credential 或免費試用額度。 常見兩大模式係大量假帳戶攞免費 credit,以及 LLMjacking:偷 AI API key 後用受害者帳戶跑模型、燒帳單 [1][4][5]。
防守重點包括收緊免費試用、設定用量上限同速率限制、掃描外洩密鑰、輪換 key、監察異常用量同開支飆升 [4][5][8]。
繼續“GPT-5.5 Instant 傳入 Microsoft 365 Copilot:企業好處同證據缺口”以獲得另一個角度和額外的引用。
開啟相關頁面對照「華為曼谷發布會:MatePad Pro Max 打頭陣,新平板、手機、手錶一次睇」交叉檢查此答案。
開啟相關頁面According to Patrick Collison, CEO of payment giant Stripe, crooks are defrauding AI firms by signing up for new accounts in order to steal tokens used to buy computing power. The problem has become so rampant, says Collison, that token thieves now account...
Cybercriminals are exploiting the booming artificial intelligence economy through a massive wave of AI token theft, forcing startups to reconsider how they acquire users. According to Patrick Collison, CEO of the payment giant Stripe, fraudsters are systema...
Over a dozen companies have suffered data theft attacks after a SaaS integration provider was breached and authentication tokens stolen. ... The TeamPCP hacking group continues its supply-chain rampage, now compromising the massively popular "LiteLLM" Pytho...
TLDR: The startup's monthly OpenAI bill was normally $400. The invoice that arrived was $67,000. Their API key had been in a public GitHub repository for 11 days. Automated bots found it within minutes of the commit. They had been running commercial AI serv...
呢個分別好重要。被偷嘅未必係密碼,而係已經通過驗證嘅存取權。SpyCloud 指,佢哋喺 2025 年重新截獲 1,810 萬個外洩 API key 同 token,並形容攻擊重心正由單純偷 username/password,轉向偷 API key、session token 同 automation credential 呢類「已登入」存取權 [6]。
大致有兩條路線:一係濫用免費 credit,一係偷 API key。
部分攻擊者會自動化開大量新帳戶,專門攞免費試用、promotion credit 或新客算力。Fortune 報道,Stripe 行政總裁 Patrick Collison 曾表示,在某些 AI 公司情境下,token 竊賊已佔新客註冊一大部分,比例達每 6 個新註冊就有 1 個 [1]。
呢個數字唔應該當成全行業標準,但足以說明:AI onboarding funnel 已經唔再只係增長渠道,亦變成詐騙入口。
有 LLMjacking 案例提到,一間初創平時每月 OpenAI 帳單大約 400 美元,但因為 API key 曾經放咗喺公開 GitHub repository 11 日,結果收到 67,000 美元帳單;報道指自動化 bot 喺 commit 後數分鐘內已經發現條 key [4]。另一份防禦指南亦指出,呢類攻擊已由機會式偷 key,演變成更有組織、針對 AI 供應商同雲端 AI 服務嘅濫用模式 [
5]。
AI 初創通常要快:自助註冊、即開即用、免費 credit、即時 demo、即時 API access。呢啲設計對增長有利,但當每次請求背後都可能係昂貴 GPU-backed inference,就同時打開咗 fraud surface [1][
2]。
另一個問題係密鑰外洩。CSO 引述 Wiz 研究指,Forbes AI 50 公司之中有 65% 被發現有已驗證 secret leak,包括 API key 同 access token 曾外洩到 GitHub [8]。呢唔代表每次外洩都一定變成 token theft,但反映快速開發環境入面,值錢 credential 好容易走漏。
AI fraud 嘅經濟效應亦唔同。普通假 SaaS 帳戶可能只係令指標變靚或浪費少量資源;但一個假 AI 帳戶,或者一條被偷 API key,可以即刻燒模型供應商 credit、GPU 推理成本同雲端開支 [1][
4][
5]。
Token theft 難偵測,因為攻擊者好多時用緊有效 key、有效 session,或者有效新帳戶。關於 token theft 嘅安全簡報警告,偷來嘅 session cookie、OAuth token 等憑證,可以令攻擊者繞過認證控制,冒充合法用戶 [11]。
所以,AI 公司要睇嘅唔只係「有冇成功登入」,而係行為有冇突然變樣。例如:
冇一粒銀彈可以解決,因為 token theft 夾喺 fraud、身份安全同雲端成本控制之間。比較可靠嘅做法係三樣一齊做。
免費 credit 唔應該只當 marketing cost,而要當成可被濫用嘅 spend exposure。AI 團隊可以考慮較細預設 trial 額度、分階段解鎖 credit、按帳戶或 API key 設 quota、rate limit,以及喺用量突然飆升時即時告警 [1][
4][
5]。
團隊應該假設 API key 遲早會外洩,除非開發流程主動阻止。repository 同 CI/CD secret scanning、定期 key rotation、最小權限 credential、快速 revoke 已曝光 key,都應該係基本動作;尤其喺 AI 公司 GitHub credential 外洩已有報道嘅背景下更係如此 [4][
8]。
只檢查註冊資料嘅 fraud system,可能捉唔到被偷 API key;只睇 login event 嘅 security system,又可能漏咗免費 credit farming。AI 平台需要將帳戶年齡、credit 消耗、API volume、模型選擇、spend velocity 放埋一齊分析,先有機會喺帳單爆炸前截停濫用 [1][
4][
5]。
當一間初創開始將 token 視為金融資產,而唔只係技術憑證,spend cap、異常偵測、key lifecycle management 就唔再係後勤安全工作,而係產品基建核心。
AI token theft 本質上係針對 AI 平台「用量表」嘅詐騙。被偷走嘅可以係 API key、session token、OAuth token 或免費試用 credit,但最終被套現嘅係付費 compute [3][
6]。
LLMjacking attacks cost victim organizations up to $100,000 per day in stolen AI compute, and the frequency is accelerating. Since the Sysdig Threat Research Team coined the term in May 2024, what began as opportunistic credential theft has evolved into a s...
New Report Highlights Surge in Exposed API Keys, Session Tokens, and Machine Identities, and more. ... “We’re witnessing a structural shift in how identity is exploited,” said Trevor Hilligoss, Chief Intelligence Officer at SpyCloud. “Attackers are no longe...
Experts say the leaks highlight how fast-growing AI firms may be prioritizing innovation over basic DevSecOps hygiene, leaving valuable intellectual property and data at risk. Nearly two-thirds of the world’s top private AI companies have exposed API keys a...
Token theft is now one of the fastest-growing threat vectors in enterprise security. Attackers steal digital passports – session cookies, OAuth tokens, SAML assertions – to bypass authentication, including multifactor authentication (MFA). Stolen tokens all...