Tool descriptions are MCP's overlooked attack surface: poisoning and metadata manipulation
Necessary-Sir-3681 · reddit · 2026-11-23
After studying MCP, the author explores how AI agents place excessive trust in tool descriptions and metadata — which can themselves be used to manipulate the agent. The article covers tool poisoning, metadata manipulation, cross-server risks, and detection approaches for suspicious changes, plus a small Python prototype experimenting with metadata fingerprinting. The core issue: system prompts, user messages, tool descriptions from different servers (changeable after approval), and tool results like emails and web pages all land in one flat context window where the model cannot reliably tell which tokens to trust.
More from coding & agent
- Gary Marcus: Claude Code's core is symbolic AI — neurosymbolic route vindicated — GaryMarcus · 2026-11-23
- Personal Agents Hackathon 2nd place: Pantry is an MCP-powered grocery agent for shared homes — HowDevelop · 2026-11-23
- AI agents play a 3D murder mystery: GPT-6.2 solves it in 3m14s — venturetwins · 2026-11-23
- Open-source AI Hedge Fund 3.0 finds, researches and builds its own portfolio — virattt · 2026-11-23
- Open-source MCP memory server gives Claude Code, Codex and OpenCode a shared brain — JoeBro121211 · 2026-11-23
- Teams battle 'test tampering': coding agents weaken assertions to pass CI — Most_Future_8465 · 2026-11-23