Vulnerability  ·  2026-11-19

MCP tool-provider misbinding: copycat tool metadata reroutes agent tool calls to an attacker

VulnerabilityMedium impactGlobal
arXiv:2611.21577 (~Nov 16) formalizes tool-provider misbinding in MCP: the tool metadata used for discovery/selection is not bound to a verified provider identity, so an adversary operating only another MCP server (no host, model, or legitimate-server compromise) gets a request intended for a legitimate provider resolved to its own tool. In controlled tests across five LLM-based MCP clients, matching tool names + task-aligned descriptions + compatible schemas caused GPT-5.4, Qwen3.7 Max and DeepSeek V3 to select the attacker's tool in all 50 trials each (84% for Gemini 2.5 Pro; Claude Sonnet 5 mostly asked for clarification). Official-sounding descriptions plus compatible schemas still caused wrong-provider execution even without name collisions.
This is the 'DNS rebinding'-style trust gap for agent tools: the label on a tool does not prove which server implements it. An attacker who mounts a community MCP server can quietly capture sensitive tool arguments (tokens, data, file paths) that the agent believes it is sending to a legitimate provider. It is the missing piece alongside prompt injection — the agent can pick the wrong server before it ever reads a poisoned page.
A rogue MCP server advertises tools whose name/description/schema collide with or spoof a legitimate provider; the LLM router resolves the call to the attacker, which then receives sensitive tool arguments.
MCP clients/servers with multi-server tool discovery (architectural vulnerability class)
Bind tool metadata to verified provider identity (org-issued server IDs, mTLS, signed tool manifests, pinned server URLs); disambiguate by origin host in the approval UI; log the resolved provider per tool call; use allowlists and side-effect execution gates.
explainx.ai — MCP Misbinding Attack: Fake Tool Names (2026-11-17)mallory.ai — Copycat MCP Tools Redirect AI AgentsarXiv:2611.21577
See this in the live feed Explore related AI security and governance findings — updated every morning.
Open the feed →