TechFlow News: On March 12, GoPlus Security (@GoPlusSecurity) announced that its security engine, AgentGuard, conducted a full-scan of the top 100 most-downloaded AI Agent skills on ClawHub. The results showed that 21% of these skills were directly blocked (high-risk), 17% were flagged as warnings, and 62% passed the scan.
High-risk skills primarily involve headless browser automation (e.g., agent-browser), email and communication control (e.g., agentmail, WhatsApp Business), high-privilege enterprise SaaS APIs (e.g., Google Workspace Admin, Feishu), and deep-search scraping tools—posing risks such as SSRF attacks, prompt injection, and data tampering.
The report recommends mandatorily introducing a “human-in-the-loop” (HITL) confirmation mechanism for high-risk skills and strictly filtering content returned by search-related skills to mitigate indirect prompt injection risks.
