Hugging Face 惡意軟體利用虛假 OpenAI 模型誘導用戶
惡意 hugging face 倉庫冒充 openai 的隱私過濾器，通過受損的 loader.py 文件向 windows 機器傳送竊取憑據的惡意軟體，隨後被移除。
attackers uploaded a repository to hugging face named 'open-oss/privacy-filter' that closely copied openai's legitimate model card but included a malicious loader.py file designed to execute powershell commands on windows hosts. hiddenlayer found that the script disabled ssl verification, fetched a remote payload from jsonkeeper.com, and installed a rust-based infostealer targeting discord, cryptocurrency wallets, and browser credentials. the repository artificially inflated its popularity to reach the platform's trending list, and six other repositories with identical infrastructure were identified. this incident highlights the risk of public ai registries serving as vectors for supply chain infiltration, as traditional software composition analysis tools fail to detect malicious logic within ai model setup scripts.