MCP data exfiltration paths

Data can leave through tool arguments, tool results, sampling context, elicitation destinations or server egress.

01 / The operating case

What changes in a real workflow?

A document agent reads an internal file and then calls a “diagnostic” MCP tool with the file contents as an argument.

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

Classify source data and outbound destinations, restrict egress, and authorize the upload as a distinct action tied to the task.

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 read grant silently becomes an export grant because the outgoing tool is labeled diagnostic.

Evidence to retain

Correlate read source, argument digest, destination, decision and network effect; avoid duplicating the sensitive payload.

04 / Canonical scope

Why this reference stands alone

Tool poisoning owns one instruction entry point; this page maps all relevant outbound data routes.

Protocol or attack trace

Outbound path inventory

Sequence

Internal document content moves through sampling context, then becomes an argument to a diagnostic upload tool.

Negative test

Deny upload to unapproved destination even though original document read was authorized.

Primary references

Read the underlying material

Architecture discussion

Choose one consequential action and make its boundary explicit.

Request a Conversation