Skip to the content
AI and Cybersecurity Learning and Practice Center Test automation Open the Learning Center

WeaknessASC-2026-0011Major weakness

MCP tool description poisoning

HighMitigated

An MCP server hides instructions for the model inside tool names, descriptions or parameter docs, for example telling it to read secrets or to change how other tools behave.

How to detect it

  • Tool descriptions contain instructions aimed at the model, secrecy phrases or references to sensitive files
  • Invisible Unicode in tool metadata
  • A tool description mentions other tools (shadowing)
  • Descriptions change between versions

How to resolve it

  • Pin MCP servers by version and approve the tool list fingerprint
  • Review and diff tool descriptions on every upgrade
  • Strip invisible characters from tool metadata before it reaches the model
  • Use clients that show users the full tool description

How to investigate

  1. Fetch tools/list from the server as it was during the incident
  2. Diff it against the approved tool list and inspect changed descriptions
  3. Search the agent transcript for tool calls that match the hidden instructions

Sources

Open it in the Learning Center's vulnerability database