Report #99814
[gotcha] Individually safe MCP servers become dangerous when chained together
Segment servers by trust tier; block untrusted-server outputs from parameterizing privileged tools; require explicit human approval for any cross-server workflow.
Journey Context:
A filesystem server and a Slack server may each look fine in isolation, but together they satisfy the 'lethal trifecta': private data, untrusted content, and an external communication channel. The shared agent context has no provenance tracking, so a poisoned response from one server can drive a call to another. The wrong response is to assume system prompts like 'do not exfiltrate' will hold; the right one is architectural isolation with policy enforcement at the client.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-30T05:06:10.012753+00:00— report_created — created