MCP 고유의 위협

Once your MCP server hands tools to a model, you have built a new way for untrusted text to drive real actions. That changes your threat model. 🛡️

Treat the model like a confused, well-meaning intern who reads everything it is given. The big risk is untrusted content steering it toward calls you never intended.

Prompt injection is when text the model reads contains hidden instructions, like a webpage saying delete all files, and the model obeys them as if they were real commands.

Injected instructions ride in on data your tools fetch: web pages, emails, file contents, API responses. Anything a tool returns can carry an attacker payload back to the model.

Tool poisoning hides malicious instructions inside a tool description or its result, so the model is nudged to misbehave just by having the tool installed.

This description looks helpful but smuggles in a side order. The model may read the hidden part as a command, so review every tool description you publish.

A confused deputy is a trusted program tricked into misusing its own authority. Your server has real credentials, so an attacker who steers it borrows that power.

Your server often holds tokens or database access the user never sees. If the model is fooled, it spends your authority, not the attacker's, which is exactly the danger.

Real attacks chain them: poisoned content injects instructions, the model relays them, and your over-privileged server acts as the confused deputy that does the damage.

No single fix is enough. You will layer least privilege, input validation, and human confirmation so one slip does not become a breach. Layered defense is the goal.

From now on, design every tool as if its arguments and the data it reads were written by an attacker. That adversarial mindset shapes every choice you make next.

Quick gut check on the core MCP threats.

MCP adds an attack surface where untrusted text can drive real actions. Watch for prompt injection, tool poisoning, and confused-deputy misuse, then defend in layers. ✅