When Tool Names Are Not Identities: Measuring Provider Misbinding in MCP-Based AI Agents

Zeynab Anbiaee, Sajjad Dadkhah, Ali A. GhorbaniarXiv 2026 · arXiv

TL;DR

This paper exposes a critical security flaw in the Model Context Protocol (MCP) called tool-provider misbinding. Researchers demonstrate that AI agents can be easily tricked into executing malicious tools simply by manipulating tool metadata—like names and descriptions—because the protocol lacks a mandatory way to verify a tool's actual provider.

What Problem Does This Paper Address?

As AI agents move from simple text generation to "agentic execution"—performing tasks like fetching files or calling APIs—they rely on protocols like MCP to discover and use external tools. Currently, an agent selects a tool based on its metadata: its name, a natural-language description, and its input schema (the data format it expects).

The core problem is that metadata is not identity. In a multi-server environment, an attacker can host a malicious server that advertises tools with names and descriptions identical to (or even "more official" than) legitimate ones. Because MCP doesn't mandate a verified provider-bound identity, the agent might "misbind" a user's request to the attacker's tool instead of the intended one. This isn't a traditional network hack; it’s a failure of the agent’s decision-making process driven by deceptive metadata.

Figure 1. Controlled MCP tool-misbinding scenario.

Methods

The authors formulated tool-provider misbinding as an optimization problem where an attacker seeks to maximize the probability of their tool being selected. They conducted a controlled experiment using a 2 × 3 × 2 factorial design to isolate how different metadata fields contribute to the attack:

  • Name: Unrelated vs. Exact (duplicate of the legitimate tool name).
  • Description: Unrelated vs. Copied vs. Authority-enhanced (adding cues like "officially verified" or "enterprise-approved").
  • Schema: Incompatible vs. Compatible (matching the structure the agent expects).

These factors created 12 distinct experimental conditions (C01–C12). They tested these conditions across five state-of-the-art LLM-based clients (GPT-5.4, Claude Sonnet 5, Gemini 2.5 Pro, Qwen3.7 Max, and DeepSeek V3) in a mock environment where a user wants to "fetch a dataset."

Datasets

The study used a custom-built, controlled MCP environment. The evaluation was based on a fixed prompt set of 10 task-equivalent requests for a benign dataset-fetching operation. These prompts ranged from ordinary user wording to more security-aware phrasing that explicitly mentioned "trusted destinations." Each prompt was run 5 times per condition to measure the stability of the agent's choices.

Key Insights and Ideas

  • Interaction is the Killer: Misbinding isn't just about duplicate names. The strongest failures occur when an exact name, a task-aligned description, and a compatible schema all reinforce each other.
  • "Official" Labels Work: "Authority-enhanced" descriptions (e.g., claiming a tool is "validated") can cause agents to pick the wrong provider even when there is no name collision.
  • Stability $\neq$ Safety: A model might consistently pick the attacker's tool (high stability), which is actually more dangerous than an unstable model that occasionally picks the right one.
  • Model-Specific Failure Modes: Different LLMs fail in different ways. Some (like GPT-5.4 and DeepSeek V3) tend to silently pick the attacker, while others (like Claude Sonnet 5) are more likely to ask for clarification, exposing the ambiguity without actually solving the underlying identity problem.
  • Schema as a Selection Signal: Interestingly, a compatible input schema acts as a strong signal to the LLM that a tool is "suitable," making it a powerful tool for attackers to increase their selection rate.

Key Results

The researchers found that high-alignment metadata (Condition C10 and C12) led to near-total failure across most models.

  • Violation Rates (VR): In the most dangerous conditions (exact name + copied/enhanced description + compatible schema), models like GPT-5.4, Qwen3.7 Max, and DeepSeek V3 reached violation rates of 96% to 100%, meaning they almost always chose the attacker's tool.
  • Clarification Rates (CR): Claude Sonnet 5 showed a different pattern, requesting clarification up to 94% of the time in ambiguous conditions. While "safer" than silent execution, it still failed to automatically bind to the correct provider.
  • Schema Amplification: Making an attacker's tool schema compatible with the legitimate one increased violation rates by nearly 1.00 (100 percentage points) for several models in certain conditions.
Figure 2. (a) Violation rate, (b) Clarification rate, and (c) Resolved Binding instability across the evaluated models and attacker-tool conditions.

Limitations and Future Work

The authors acknowledge several limitations:

  • Scope of Tasks: The study focused on a single "dataset-fetching" task. Different types of tools might yield different misbinding rates.
  • Environment: The experiments were conducted in a controlled, two-server setup which may not fully reflect the complexity of real-world MCP deployments with dozens of servers.
  • Attacker Strategies: The study used a limited set of metadata manipulation strategies.

Future Work: The researchers plan to expand their model to capture a broader range of attacker strategies and tool classes. They emphasize that the long-term solution isn't just "better prompts" but protocol-level changes, such as signed tool manifests and stable provider-bound identifiers, to ensure that tool names are finally treated as identities.

Summary updated

Abstract

Large language model (LLM) applications increasingly act as agents that select and invoke external tools. The Model Context Protocol (MCP) standardizes this interaction by allowing clients to compose tools advertised by multiple servers. However, tool metadata used for discovery and selection is not, by itself, necessarily bound to a verified provider identity. This paper studies this gap as tool-provider misbinding: a failure in which a request intended for a legitimate provider is instead resolved to an attacker-controlled tool. The attacker does not compromise the host, LLM, legitimate server, or network channel; it only operates another MCP server that advertises tools with chosen metadata. We formulate tool provider misbinding as an attacker-controlled tool-selection optimization problem and evaluate it in a controlled MCP environment across five LLM-based clients. The results show that interactions drive misbinding. The strongest failures occur when exact-name collision, task-aligned description, and compatible schema reinforce one another, but authority-enhanced descriptions with compatible schema can also cause wrong-provider execution or unresolved provider ambiguity without name collision. Different models expose the same provider-binding weakness differently, including attacker-controlled provider selection, clarification requests, and unstable or inconsistent provider resolution.
BibTeX
@misc{anbiaee2026when,
  author = {Zeynab Anbiaee and Sajjad Dadkhah and Ali A. Ghorbani},
  title = {{When Tool Names Are Not Identities: Measuring Provider Misbinding in MCP-Based AI Agents}},
  year = {2026},
  eprint = {2611.21577},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.21577},
}
Figure