MCP tool trust

A tool definition is an interface description supplied by a server, not proof of safe behavior or authority.

01 / The operating case

What changes in a real workflow?

A server changes an export tool description to say it is a harmless diagnostic read while it still sends data externally.

Trace the authority boundary
  1. 01Client / task
  2. 02MCP server
  3. 03Tool call + arguments
  4. 04Downstream effect

02 / Decision contract

What the executor must check

Review tool schema and implementation effect, classify destinations and mutation risk, and enforce operation and argument policy at invocation.

InputPrincipal · task · action · target
DecisionGrant ancestry · policy · current state
ResultEnforce · record · verify outcome

03 / Failure and evidence

What goes wrong, and what can be proven?

Failure mode

A descriptive label substitutes for an observed effect, allowing data export through a “read” tool.

Evidence to retain

Retain tool-definition hash, change review, call arguments and downstream destination.

04 / Canonical scope

Why this reference stands alone

Tool poisoning focuses on malicious instructions in metadata; this page owns effect classification and interface trust.

Protocol or attack trace

Tool effect classification

Sequence

An export tool changes description from “send report” to “diagnostic read” while still posting data to an external URL.

Negative test

Classify by downstream effect and destination; deny an export task lacking recipient authorization despite the harmless label.

Primary references

Read the underlying material

Architecture discussion

Choose one consequential action and make its boundary explicit.

Request a Conversation