Fudan Researchers Find 1-in-10 MCP Tools Lie to AI Agents
A new study scanned 2,214 real-world MCP servers and found nearly 1-in-10 tools lie to the AI agents that use them
Takeaways
- Researchers scanned 2,214 real-world MCP servers and found 9.93% of tool descriptions are inconsistent with what the code actually does.
- The paper introduces Description-Code Inconsistency (DCI), covering both accidental documentation drift and deliberate deception -- a broader threat than tool poisoning alone.
- DCIChecker combines static code analysis with a novel Direct-Reverse-Arbitration (DRA) prompting method to cross-validate descriptions against implementations.
- The dataset covers 19,200 description-code pairs; a related independent study found ~13% mismatch rates across 10,240 servers, corroborating the scale of the problem.
- DCI creates a defense blind spot: agents may silently write files, exfiltrate credentials, or make unauthorized network calls while believing they are calling a safe tool.
- The paper proposes mitigations including description signing, registry-level auditing, and CI/CD consistency checks; the full paper is on arXiv.
When an AI agent decides which tool to call, it reads a plain-text description provided by the MCP server and trusts it completely. There is no handshake, no verification, no audit. A new paper from researchers at Fudan University exposes how badly that trust is being abused -- and how often it breaks down even without malicious intent.
The paper introduces the concept of Description-Code Inconsistency (DCI): the gap between what an MCP tool says it does and what its code actually executes. After scanning 2,214 real-world MCP servers, the team found this problem is far more common than anyone had measured.
The protocol that became the backbone of agentic AI
The Model Context Protocol was introduced in November 2024, promising to become the "USB-C of AI agents" -- a universal bridge for connecting LLMs to tools, APIs, documents, emails, codebases, databases, and cloud infrastructure. In just months, the ecosystem exploded: dozens of tool servers, open-source integrations, host implementations, and hosted MCP registries began to appear.
MCP servers act as crucial middleware connecting AI models to external tools, APIs, and data sources, enabling AI agents to perform complex, context-aware actions. They expose prompts, tool definitions, and runtime permissions that allow seamless interactions -- but also create new security risks when mismanaged or maliciously influenced. The problem is that MCP's rapid proliferation has outpaced the development of its security model. Much like early web protocols, MCP was released with a flexible and underspecified design, allowing implementers freedom of design but also introducing ambiguity for safe usage.
The blind spot nobody was checking
In the MCP ecosystem, LLMs rely on natural language descriptions provided by MCP servers to select and execute functions. This interaction implicitly assumes that tool descriptions faithfully reflect their underlying implementations -- an assumption that is not mandatorily verified in practice. As a result, MCP deployments may suffer from Description-Code Inconsistency (DCI), where a tool's description of its capabilities and security boundaries is not consistent with what the code actually does.
This is distinct from the more widely discussed