Agent and MCP security: confused deputies, tool poisoning and least privilege
The attack surface of tool-using agents and MCP servers: confused-deputy attacks, indirect injection through tool outputs, malicious or poisoned tool descriptions, over-broad credentials, rug-pull updates, data exfiltration paths, and a defence checklist with authorisation design.
17 min·Verified 2026-11Pro+50 XP
An LLM that only produces text can embarrass you. An agent with tools can act. Security for agents is mostly about limiting what a manipulated model can do (threat model, excessive agency).
Pro content
Keep reading with Inference Academy Pro
Unlock the rest of this lesson, plus every advanced lesson, hands-on lab, the larger playground models and completion certificates.