A new security vulnerability has been identified in how coding agents handle skills provided by MCP servers. While approving a server's skills pins file hashes, this only confirms the files haven't changed, not that they are safe. The system allows skills to be added to the agent's context, but the safety of these instructions is not verified until the skill is loaded, and instructions pointing to external web addresses are not hashed at all. The recommended solution is to treat the approval as a lockfile entry, requiring users to review each skill before allowing it and maintain a personal record of approved skills. AI
IMPACT Highlights a critical security flaw in AI-powered coding tools, necessitating user vigilance and improved vetting processes for external skill integrations.
RANK_REASON The article discusses a security vulnerability in a specific tool (coding agents with MCP server skills) rather than a new model release or fundamental research.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →