Analysis
MCP Tool Poisoning: What It Is and How to Stop It
First reported · Discovered eseospace.com
Page published
Earliest dated coverage: 28 Oct 2026 · First observed: 31 Oct 2026 · Latest dated coverage: 28 Oct 2026
Coverage timeline
Single-source analysis — one report is available.
Why it matters
MCP tool poisoning exposes a trust boundary between server-provided metadata and agent instructions, potentially allowing connected tools to redirect behavior or extract sensitive data.
eSEOspace’s explainer describes MCP tool poisoning, where malicious instructions embedded in tool descriptions, schemas, or metadata influence an AI agent, potentially before any tool executes. The article reviews Invariant Labs’ previously published demonstration involving sensitive-file exfiltration and recommends trusted servers, visible descriptions, version pinning, and change alerts.