OpenClaw：恶意技能入侵 ClawHub，AI 代理及加密钱包受威胁

Core Insights - ClawHub experienced a significant influx of malicious skills, with 1,184 instances reported between late January and mid-February 2026, accounting for 36.8% of total uploads [1] Group 1: Malicious Skills and Attacks - A single attacker uploaded 677 packages, with 12 accounts involved in the malicious activity [1] - The malicious skills disguised themselves as cryptocurrency trading bots, YouTube summarizers, and wallet trackers, achieving thousands of downloads [1] - The package "What Would Elon Do" contained 9 vulnerabilities, including 2 critical ones, and was boosted by 4,000 fake downloads [1] Group 2: Attack Methods and Impact - Attack methods included using SKILLmd documents to induce the execution of curl | bash commands to install malicious programs, with macOS versions being a variant of Atomic Stealer and Windows versions being VMProtect stealers [1] - The attacks involved injecting prompts to manipulate AI agents, leading to the theft of browser passwords, over 60 cryptocurrency wallets, SSH keys, Telegram sessions, Keychain credentials, .env files, and OpenClaw configurations [1] - More than 135,000 instances were affected, distributed across 82 countries [1]