One in Four AI Agent Skills Has a Security Hole, and Your Agent Runs Them With Your Permissions
中文摘要
研究发现，AI智能体生态中25%的技能存在安全漏洞。由于智能体使用用户权限运行，这些漏洞构成了极高的安全风险。
English Summary
Researchers found 25% of AI agent skills have security vulnerabilities. Since agents execute them with user permissions, these flaws pose high-stakes security risks.
Original Excerpt
Researchers audited 42,000 skills across the new agent ecosystem. The findings look a lot like npm’s early years, only with higher stakes. Continue reading on Generative AI »