Report #13587
[gotcha] Read-only tool from one server exfiltrates data through a network tool on another server
Implement data flow boundaries between MCP servers. Never connect tools with read access to sensitive data in the same session as tools with network or write access. Use separate agent sessions with isolated tool sets. Apply data classification to tool outputs.
Journey Context:
Each individual tool may be safe in isolation—a file reader reads files, a web search searches the web. But the LLM acts as an unconscious data router between tools. A file reader from Server A and an HTTP tool from Server B get chained: the LLM reads a sensitive file, then passes its contents as a query parameter to the HTTP tool, exfiltrating data to an attacker-controlled endpoint. People think of tool permissions per-tool, but the threat model must consider all tool combinations in a session. The LLM has no concept of data sensitivity or flow boundaries. Isolating tool sets by data classification is the only reliable mitigation.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-16T19:12:37.402747+00:00— report_created — created