MCP, the Model Context Protocol, is the standard most agents now use to plug in tools: a Gmail server, a calendar server, a file server, a smart-home server. Each MCP server you add is either code running on your machine or a remote service, and your agent will act on what it says. That makes MCP servers part of your attack surface, in the same way that agent skills are.
This page covers the risks that matter for personal use. Developers building servers should read the full MCP security best practices in the specification.
How big is the MCP ecosystem?
As of December 2025, Anthropic reported "more than 10,000 active public MCP servers" and "97M+ monthly SDK downloads across Python and TypeScript" (Anthropic). Google Home opened early-access MCP control to third-party agents including Claude and OpenClaw in September 2026, with doors, alarms and garage doors kept off-limits. Grok Bot uses connectors and MCP where available. Background on how connectors work is on getpersonalaiagents.com.
- Active public MCP servers (Anthropic, Dec 2025)
- 10,000+
- Monthly MCP SDK downloads (Anthropic, Dec 2025)
- 97M+
- Tool poisoning first described by Invariant Labs
- Apr 2025
What are the main MCP security risks?
Tool poisoning
In April 2025, Invariant Labs described tool poisoning attacks: "malicious instructions are embedded within MCP tool descriptions that are invisible to users but visible to AI models." Your client may show a one-line summary of a tool while the model reads a much longer description. It is a form of indirect prompt injection, except the attacker is the tool author.
Rug pulls
Invariant also described the "rug pull": "a malicious server can change the tool description after the client has already approved it." You approve a harmless tool, and a later update changes what it tells the model. This mirrors malicious package updates in software registries.
Tool shadowing across servers
A malicious server does not even need to be called. Invariant showed that one server's tool description can change how the agent uses a different, trusted server's tools, in its test redirecting where emails were sent. With several servers connected, one bad one can contaminate the others.
Malicious or compromised local servers
Many MCP servers run locally. The MCP specification warns that local servers "may have direct access to the user's system" and lists the risks of untrusted ones as arbitrary code execution, no visibility into what runs, obfuscated commands and data loss. It requires that clients offering one-click local setup "show the exact command that will be executed, without truncation" and get explicit approval, and recommends sandboxing servers with minimal default privileges.
Over-broad connector scopes
The quieter risk is a legitimate connector that asks for far more than it needs. The spec's section on scope minimisation describes the problem: a token "carrying broad scopes" granted up front means a single leak or mistake exposes everything. It recommends starting with a minimal scope set and asking for more only when a privileged action is first attempted.
Messaging servers as an injection channel
Invariant's WhatsApp MCP research showed an injected message, or a malicious companion server, making an agent exfiltrate chat history and contacts. Connecting a messaging server lets anyone who can message you reach your agent.
- Malicious or careless MCP server
- Hidden instructions or broad token
- Agent with other connectors
- Data sent out or wrong action taken
Why does mixing servers make it worse?
Every server you add can increase all three legs of the lethal trifecta: more private data, more untrusted content, more ways to send data out. Simon Willison's warning applies directly: "once you start mixing and matching tools yourself there's nothing those vendors can do to protect you." A hosted agent's built-in connectors are tested together. Your own combination of community servers is not. A practical rule: avoid connecting a server that reads untrusted content, such as a web fetcher or a messaging bridge, to the same agent that holds your email or files, unless every outgoing action needs your approval.
How do hosted agents handle MCP risk?
Hosted agents choose which connectors they offer and add their own controls around them, which removes some of the risk but not all. Examples from vendor documentation:
| Agent | Relevant control | Source |
|---|---|---|
| Meta Muse | Approvals tied to a specific connector and use; agent sees only surrogate tokens | Meta |
| Claude Duet | Never sees sign-in tokens for connected apps; action check runs outside the model | Anthropic |
| Grok Bot | Approval depends on tool, risk and your Auto-review rules | xAI |
| OpenClaw | You choose and run servers yourself; per-agent tool allow and deny lists | OpenClaw docs |
As of November 2026.
How do you use MCP servers more safely?
- Install fewer serversEach one is code or a service you are trusting with your agent's reach.
- Prefer official serversUse servers published by the service owner over unknown community copies with similar names.
- Read the full tool descriptionsIf your client lets you, check what the model actually sees, not just the summary.
- Pin and re-checkPin versions, and re-review tools after any update or new approval request.
- Grant minimal scopesChoose read-only or narrow scopes on the consent screen when offered.
- Keep write actions gatedRequire approval for sending, sharing, deleting and purchasing through any connector.
See least privilege for AI agents and how to revoke AI agent access. OpenClaw users can also apply per-agent tool restrictions from the OpenClaw hardening checklist.
Bottom line
MCP made it easy to give agents new abilities, and that ease is the risk. Tool poisoning, rug pulls and shadowing mean a server's text can steer your agent, and broad scopes mean a single mistake exposes a whole account. Treat every MCP server like software you are installing with your agent's permissions, because that is what it is.



