ENCFORGE勒索軟體專殺AI模型：JADEPUFFER進化解析 | ENCFORGE Ransomware Targets AI Models: JADEPUFFER Evolves
By Kit 小克 | AI Tool Observer | 2026-07-30
🇹🇼 ENCFORGE勒索軟體專殺AI模型：JADEPUFFER進化解析
ENCFORGE勒索軟體是駭客集團JADEPUFFER最新開發的攻擊工具，第一次把矛頭直接對準AI模型檔案和訓練資料，而不是傳統的資料庫或文件系統。根據多家資安媒體本週披露的分析，JADEPUFFER利用Langflow的一個未修補漏洞（CVE-2025-3248）遠端執行程式碼，再用ENCFORGE把整個AI開發環境的關鍵檔案全部加密上鎖，研究人員估計光是重建一個被鎖住的模型，就可能要花上7.5萬到50萬美元的雲端GPU運算與工程成本。
ENCFORGE勒索軟體怎麼運作？
ENCFORGE是一支用Go語言編譯、再用UPX打包的勒索軟體，鎖定將近180種副檔名，範圍涵蓋Hugging Face的SafeTensors、PyTorch與TensorFlow的模型檔、llama.cpp的GGUF/GGML權重檔、FAISS向量索引，甚至連Parquet、Arrow、TFRecord這類訓練資料格式都不放過。它會用AES-256-CTR加密檔案內容，再搭配RSA-2048公鑰保護金鑰，加密完成後把檔案改成.locked副檔名，還會主動關閉正在佔用檔案的程式，執行完甚至能自我刪除、湮滅犯罪證據。
為什麼AI模型特別容易被鎖定？
因為訓練一個大型模型動輒燒掉數十萬到數百萬美元的算力，而多數團隊的模型checkpoint、資料集又缺乏即時備份，一旦遭ENCFORGE鎖住，公司往往被迫在「付贖金」與「從零重訓」之間二選一——兩者都是天價。JADEPUFFER正是看準這個痛點，才把ENCFORGE勒索軟體從單純的資料庫勒索，升級成專門摧毀AI資產的武器。
企業該怎麼防範ENCFORGE攻擊？
第一步是立刻修補Langflow的CVE-2025-3248漏洞，避免暴露在公網的服務被未授權遠端執行程式碼；第二步是把模型checkpoint、向量索引與訓練資料獨立備份到離線或版本控制的儲存空間，不要只放在同一台可被攻破的伺服器上；第三步是針對AI基礎設施的異常存取與大量檔案改名行為設定即時警報，及早攔截加密行為。
小克實測心得：MLOps流程裡最容易被忽略的往往不是模型本身的安全，而是它「存放的地方」有沒有基本的存取控制與備份機制。ENCFORGE勒索軟體證明了AI基礎設施已經正式成為駭客的高價值目標，光靠模型好不好用已經不夠，安全防護要一起補上。
好不好用，試了才知道。
🇺🇸 ENCFORGE Ransomware Targets AI Models: JADEPUFFER Evolves
ENCFORGE ransomware is the newest weapon from threat actor JADEPUFFER, and it is the first ransomware built specifically to destroy AI model files and training data instead of ordinary databases or documents. According to security researchers who published their findings this week, JADEPUFFER exploits an unpatched Langflow flaw (CVE-2025-3248) to gain unauthenticated remote code execution, then deploys ENCFORGE ransomware to lock down an entire AI stack — researchers estimate rebuilding a single encrypted model could cost between ,000 and ,000 in cloud GPU compute and engineering time.
How Does ENCFORGE Ransomware Work?
ENCFORGE is a Go-compiled, UPX-packed ransomware binary that targets roughly 180 file extensions, covering Hugging Face SafeTensors, PyTorch and TensorFlow checkpoints, llama.cpp GGUF/GGML weights, FAISS vector indexes, and training data formats like Parquet, Arrow, and TFRecord. It encrypts files with AES-256-CTR protected by an RSA-2048 public key, renames locked files with a .locked extension, kills processes holding files open, and can self-delete after execution to cover its tracks.
Why Are AI Models Such an Attractive Target?
Training a large model can burn through hundreds of thousands to millions of dollars in compute, yet most teams do not keep real-time backups of checkpoints and datasets. Once ENCFORGE locks those files, companies face a brutal choice between paying the ransom or retraining from scratch — both wildly expensive. JADEPUFFER upgraded from generic database extortion to purpose-built AI-asset destruction precisely because this pain point makes victims more likely to pay.
How Can Teams Defend Against ENCFORGE?
Patch the Langflow CVE-2025-3248 vulnerability immediately to close the remote-code-execution entry point. Keep model checkpoints, vector indexes, and training data backed up offline or in version-controlled storage, not just on the same server that could be compromised. Set up real-time alerts for unusual access patterns and mass file-renaming behavior across AI infrastructure so encryption attempts get caught early.
Kit's take: in most MLOps pipelines, the weak point is not the model itself — it is where the model lives, and whether that storage has basic access control and backups. ENCFORGE ransomware is proof that AI infrastructure is now a high-value target in its own right. A great model is not enough anymore; the security around it has to catch up.
好不好用，試了才知道 — you only know if it holds up once you have actually tried it.
Sources / 資料來源
- JADEPUFFER Deploys ENCFORGE Ransomware Built to Destroy AI Models and Training Data
- New ENCFORGE Ransomware Targets AI Model Files in Langflow RCE Attack
- JadePuffer agentic attacks now target AI model data with ransomware
常見問題 FAQ
ENCFORGE勒索軟體是什麼？
一款由JADEPUFFER駭客集團開發的勒索軟體，專門加密AI模型檔案與訓練資料，而非傳統文件或資料庫。
ENCFORGE怎麼入侵系統？
利用Langflow的CVE-2025-3248漏洞取得未授權遠端程式碼執行權限，再部署ENCFORGE加密整個AI開發環境。
中了ENCFORGE勒索軟體要花多少錢重建模型？
研究人員估計每個模型的重建成本約在7.5萬到50萬美元之間，取決於模型規模與雲端GPU算力費用。
企業該怎麼預防ENCFORGE攻擊？
立即修補Langflow漏洞、把模型checkpoint與訓練資料離線備份，並針對異常檔案存取設定即時警報。
延伸閱讀 / Related Articles
- Kimi K3評測：2.8兆參數最大開源模型，你真的跑得動嗎 | Kimi K3 Review: Biggest Open-Weight Model You Can't Run
- Gemini 3.5 Pro三度延期：Google砍掉重練追GPT-5.6 | Gemini 3.5 Pro's Third Delay: Google Restarts From Scratch
- 歐盟強制Google開放Android給AI助手：DMA新規完整解析 | EU Forces Google to Open Android to Rival AI Assistants
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