Finding safe MCP servers: permissions, prompt injection, vetting
How MCP servers can go wrong, from poisoned tool descriptions to over-broad tokens, and a short checklist for vetting one before you connect it.
Updated 2026-11-19. "(source N)" refers to the numbered list at the end.
Why this matters
An MCP server is code that sits between an AI model and your accounts. A good one saves hours. A bad one can read your files, leak your data, or take actions you never asked for.
The MCP spec is clear about this. It says tools amount to running arbitrary code, that a server's own description of its tools should not be trusted unless the server is, and that apps must ask the user before calling any tool (source 2). Anthropic, which runs a directory of connectors, says it reviews listings against its criteria but does not security-audit the servers (source 5).
Threat 1: prompt injection
Prompt injection is when text the model reads contains instructions that hijack it. It is the top risk on OWASP's list for AI apps (source 7). The text can arrive through an email, a web page, a support ticket or a tool result.
The UK's National Cyber Security Centre warns that this is not like SQL injection, which can be fully fixed. Language models cannot reliably tell instructions apart from data, so the risk can be reduced but perhaps never removed (source 8). Plan on the basis that some attacks will get through.
Threat 2: poisoned tools
Researchers at Invariant Labs showed a sneakier version aimed at MCP itself (source 3). A server's tool description, which the model reads but the user usually does not, can hide instructions.
- Tool poisoning: a tool that claims to add two numbers quietly tells the model to read your SSH key and pass it along as a hidden argument.
- Rug pull: a server you approved later changes its tool descriptions to malicious ones.
- Shadowing: a bad server's description changes how the model uses a different, trusted tool, for example by redirecting emails sent through your real email tool.
Threat 3: the lethal trifecta
Developer Simon Willison names the most dangerous combination. An agent becomes easy to exploit when it has all three of these: access to private data, exposure to content an attacker can control, and a way to send data out (source 4). With all three, one malicious email can lead to your data being posted to a stranger's server.
Every MCP server you add may bring one of these. A web-fetch server brings untrusted content. A CRM server brings private data. An email server brings a way out. Check the combined set, not each server alone.
Threat 4: too much access
The MCP security guide warns against tokens with broad scopes like full access to every file or every database. If such a token leaks, the attacker gets everything at once (source 1). It recommends starting with minimal, read-only scopes and asking for more only when a specific action needs it.
Local servers carry their own risk. They run with the same rights as your user account. The guide asks apps to show the exact start command before installing one, and suggests running local servers in a sandbox with limited file and network access (source 1).
A vetting checklist
Before you connect a server, work through this list.
- Who made it? Prefer the official server from the vendor whose service it connects to, or one you wrote yourself (source 5).
- Where did you find it? The official MCP Registry is a public catalogue, built in the open on GitHub (source 6). Being listed means it can be found, not that it is safe.
- Is the code public and recently maintained? Read the tool list and every tool description yourself.
- What scopes does it ask for? Reject servers that want full access when read-only would do.
- Can you pin a version? Pinning protects against rug pulls, which Invariant recommends (source 3).
- What would the trifecta look like with your other servers? If it completes it, split the work across separate agents.
- Where do secrets live? Keep API keys in environment variables, not in shared config files (source 9).
Safe habits once connected
None of this means you should avoid MCP. It means treating each server the way you would treat a new browser extension with access to your bank: useful, but only from a source you trust and with no more access than it needs.
- Start read-only. Add write tools one at a time.
- Keep approval prompts on for any tool that sends, pays, deletes or publishes.
- Log tool calls. The NCSC suggests watching for failed or odd calls, which can be a sign someone is probing (source 8).
- Review your connected servers every quarter and remove ones you no longer use.
- Do not rely on any product that claims to stop prompt injection completely. Both Willison and the NCSC are sceptical of such claims (source 4, source 8).
Sources
Written by Agents Lookup in our own words. Facts come from these pages:
- 1. Model Context Protocol: Security best practices: modelcontextprotocol.io/specification/latest/basic/security_best_practices
- 2. Model Context Protocol: Specification: modelcontextprotocol.io/specification/latest
- 3. Invariant Labs: MCP security notification, tool poisoning attacks: invariantlabs.ai/blog/mcp-security-notification-tool-poisoning-attacks
- 4. Simon Willison: The lethal trifecta for AI agents: simonwillison.net/2025/Jun/16/the-lethal-trifecta
- 5. Claude Code Docs: Security: code.claude.com/docs/en/security
- 6. GitHub: modelcontextprotocol/registry: github.com/modelcontextprotocol/registry
- 7. OWASP GenAI Security Project: LLM01 Prompt injection: genai.owasp.org/llmrisk/llm01-prompt-injection
- 8. NCSC: Prompt injection is not SQL injection: www.ncsc.gov.uk/blog-post/prompt-injection-is-not-sql-injection
- 9. Claude Code Docs: Connect Claude Code to tools via MCP: code.claude.com/docs/en/mcp