Daniel Miessler • 8/25/2025

MCPs Are Just Other People's Prompts Pointing to Other People's Code

This article critically examines the security implications of Model Context Protocols (MCPs). It argues that MCPs introduce a unique trust risk because they are essentially prompts that instruct an AI agent to go and execute other people's code. The author contrasts this with traditional third-party API usage, highlighting the dynamic and less assessable nature of trusting an AI to interpret and act on external instructions, and provides a code example of a potential malicious prompt.

0 comments

#prompt engineering #mcp #ai security