Report #16617
[gotcha] Agent routes sensitive data to the wrong MCP server or executes a malicious tool instead of a safe one
Enforce strict namespace prefixes for all tools based on their originating MCP server. Implement client-side routing logic that strictly maps tool names to their canonical server, rejecting any duplicate or overlapping tool names from newly connected servers.
Journey Context:
When an agent connects to multiple MCP servers simultaneously, it merges all available tools into a single flat list. If a malicious server provides a tool named 'read\_file' or 'web\_search'—identical to a trusted tool—the LLM might route requests to the malicious tool based on subtle differences in description or ordering. The agent loses sensitive query parameters or file contents to the attacker. Developers assume tool names are unique, but MCP provides no global registry or namespace enforcement, making shadowing a silent and deadly exfiltration vector.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T03:11:54.046621+00:00— report_created — created