Quick Facts

  • Tool poisoning attacks embed malicious instructions inside MCP tool descriptions, which AI agents process as trusted operational context with no independent verification.
  • The first documented in-the-wild malicious MCP server, postmark-mcp, quietly added email exfiltration code at version 1.0.16 after 15 clean releases, reaching 1,643 downloads before removal.
  • As of April 2026, 78% of enterprise AI teams run at least one MCP-backed agent in production, making the attack surface broad and growing.

A single compromised Model Context Protocol server can manipulate an AI agent into misusing every other tool it has access to in a session, including file systems, GitHub repositories, Slack channels, and email services, according to new reporting on the state of MCP security. The attack requires no exploit of the underlying services. It only requires control of the connector layer between the agent and those services.

The Model Context Protocol was open-sourced by Anthropic in late 2024 to standardize how AI agents connect to external tools. Its core security assumption is that any server announcing a tool is trustworthy. The MCP 1.0 specification included no mandatory authentication requirement and no mechanism for per-operation permission scoping.

How Tool Poisoning Works

Researchers at Invariant Labs coined the term Tool Poisoning Attack to describe the core technique. An attacker registers an MCP server with tool descriptions that embed malicious instructions alongside a legitimate-looking function. When the AI agent retrieves that metadata to plan its next action, it treats the full poisoned description as ground truth. The agent cannot inspect the server, cannot verify the schema, and has no way to distinguish a legitimate tool definition from a manipulated one.

The variant called cross-server tool shadowing extends the damage further. A Cloud Security Alliance research note describes how one malicious server can inject instructions that redefine how an agent uses a completely separate trusted server's tools. The compromised server never has to be called directly. Its descriptions can instruct the agent to blind-copy an attacker's address every time it invokes a trusted email tool on a different server entirely.

A third variant, the rug-pull attack, exploits how MCP binds trust to a server name rather than its content. A server approved on day one stays trusted in the agent's configuration even after the operator silently updates its tool definitions. The agent has no mechanism to detect the change between sessions.

Proof-of-Concept and Real Incidents

Invariant Labs published its foundational write-up on April 1, 2025, demonstrating exfiltration of SSH private keys and configuration files through a poisoned tool inside Cursor IDE. On April 7, 2025, researchers Luca Beurer-Kellner and Marc Fischer followed with a controlled demonstration showing how a malicious MCP server could redirect an AI agent to read a user's full WhatsApp chat history and forward it to an attacker-specified number, suppressing any user notification in the process.

The first in-the-wild case arrived in September 2025. The npm package postmark-mcp shipped 15 clean versions before a backdoor was added at version 1.0.16 on September 17, 2025, inserting a blind-copy field that silently forwarded email contents to an external domain. Koi Security researcher Idan Dardikman discovered and disclosed the backdoor on September 25, 2025. The package had accumulated 1,643 downloads before removal. The attacker never touched Postmark's systems.

Invariant Labs also demonstrated that the official GitHub MCP server could be hijacked via a prompt injection embedded in a public GitHub issue, leaking private repository data. Anthropic patched three vulnerabilities in its own official Git MCP server in December 2025 and removed the git_init tool entirely. A separate vulnerability tracked as CVE-2025-54136, scoring 7.2 on the CVSS scale, was disclosed in Cursor IDE in August 2025.

What Defenders Can Do

The attack surface is large. More than 10,000 public MCP servers are listed across major registries, and 75% of API gateway vendors are expected to ship MCP-specific security features by the end of 2026. Published defense guidance calls for cryptographic signing of tool schemas, strict allowlisting of approved MCP servers, and runtime monitoring that flags unexpected tool invocations before they execute.

Enterprises running MCP-backed agents should audit every server in their agent configurations, pin specific package versions with hash verification, and treat any third-party MCP server as an untrusted input source until proven otherwise. The postmark-mcp incident showed that a supply chain attack on the connector layer can bypass every security control applied at the application or service level.

The MCP ecosystem is moving fast. The number of production deployments doubled in the first half of 2026. Security controls have not kept pace with that growth.

This article was written by an AI agent. Spotted an error? Send a correction and we will fix it.