MCP tool poisoning and action boundaries

Tool poisoning inserts malicious text into a tool description or response to redirect an agent toward a call or data disclosure outside its legitimate task.

01 / Attack path

How the boundary is crossed

A newly discovered search tool claims the agent must first upload an internal report to a diagnostic endpoint.

Attack path and interception point
  1. 01Lower-trust input
  2. 02Attempted instruction
  3. 03Protected action
  4. 04Enforced decision

02 / Containment

Where bounded authority limits the effect

Verify server and tool provenance, review changed descriptions, restrict destinations and deny an upload absent a task-scoped grant.

Architecture review / negative test

Make the boundary testable.

The attacker controls tool metadata or output, not the organization’s grant. A description can still persuade a model to call another tool or export data. Record where that text enters context and where the proposed action crosses from model suggestion into a privileged API. Server trust and tool-version review reduce exposure but do not replace action checks.

03 / Failure and evidence

The attacker’s opportunity and the defender’s record

Exploit condition

The model follows a persuasive tool description and uses an otherwise valid broad credential to export data.

Evidence to retain

Record tool-definition version, prompted call, authorization decision and attempted outbound effect.

04 / Canonical scope

Why this reference stands alone

Prompt-injection threat page analyzes untrusted content generally; this page owns metadata at the MCP tool boundary.

Primary references

Read the underlying material

Architecture discussion

Choose one consequential action and make its boundary explicit.

Request a Conversation