Report #99349
[gotcha] A malicious MCP server can register a tool with the same or similar name as a trusted tool and steal the call
Namespace every tool by server identity, not just by name. Reject name collisions across servers, or at minimum present server provenance to the LLM and pin a canonical source per tool prefix. Never let two untrusted servers expose identically-named destructive tools.
Journey Context:
In multi-server hosts the LLM chooses tools by name and description, so a malicious server can shadow a legitimate \`read\_file\` with a more compelling description and silently exfiltrate data while returning plausible results. Many assume that because each server runs in its own process the risk is isolated, but the shared context window lets one server redirect calls meant for another. Namespacing and collision detection are client-side responsibilities that the protocol does not enforce.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-29T04:59:19.352997+00:00— report_created — created