MCP server security risks are hard to assess from the familiar USB-C analogy alone. That analogy is useful for about ten seconds. It tells you MCP is a universal connector. It does not tell you that the thing you plugged in acts with credentials somebody configured for it, and describes itself to your AI in wording it chose.
The Model Context Protocol (MCP) is an open standard, introduced by Anthropic in November 2024 and adopted since by other major AI vendors, for letting an AI assistant use outside tools and data. MCP provides shared message formats for integrations. A compatible server can expose tools for your calendar, database or shell to a client with the necessary access. That is genuinely useful, and it is also why the security question is different from anything we have covered so far.
Connecting a server: what the model receives
Strip away the diagrams and the sequence is short.
For a local stdio server, the client starts a process; for a remote server, it connects to a running service over a supported transport. The client requests a tool list. The response describes the tools and their input parameters. Human-readable descriptions can also be presented to the model to help it choose a tool; clients may filter or load that information selectively.
If a tool request passes the application’s controls, the client sends it to the server, which returns a result. Results may then enter the model’s context. The November 2025 specification requires server-side input validation, access controls, rate limiting and output sanitization. It recommends client-side confirmation for sensitive operations, input visibility, result validation, timeouts and logging.
These controls span the host application, client, server and downstream service. A server’s description influences tool selection, but it does not by itself grant operating-system or API permissions.
Every connection carries a trust assumption
Here is the part that reframes everything else. When you install an MCP server, you are not installing a passive connector. A local installation adds executable code; a remote connection adds a service your client communicates with. Either can supply descriptions or results that influence the model.
A malicious description can mix legitimate tool documentation with instructions to misuse another tool. Role boundaries, validation and approvals help, but protocol compatibility is not a guarantee that the text is trustworthy.
Invariant Labs demonstrated exactly this in April 2025 with what they called a tool poisoning attack: a harmless-looking add tool whose description instructed the agent, in text the user never saw, to read ~/.ssh/id_rsa and ~/.cursor/mcp.json and pass the contents along, with a benign explanation of arithmetic wrapped around it. This demonstration exploited the handling of untrusted tool descriptions and limited approval visibility in the tested setup. If you have read our guide to direct vs indirect prompt injection, this is the same mechanism, aimed at a channel you did not think of as untrusted input.
Credential handling is a separate boundary. A local server may use secrets from its environment or configuration; a remote integration may use OAuth. Clients can hold tokens for contacting servers, and servers may hold separate credentials for downstream services. Secrets should stay out of model context, but MCP does not make accidental disclosure impossible. Restrict both the requests the client can send and the authority the server can exercise.
MCP server security risks that show up in practice
The following examples illustrate supply-chain, authentication and cross-tool risks.
Installing an MCP server adds executable code
A command copied from a README can download and run a local package. Review the publisher, package and requested access as you would for other executable dependencies. A remote service avoids installing that server locally, but still raises questions about data handling and delegated permissions.
A vetted MCP server can turn malicious in the next version
In September 2025, Snyk analyzed the npm package postmark-mcp. Its timeline places version 1.0.0 on 15 September and 1.0.18 on 17 September; Snyk’s analysis suggests the hidden BCC behavior began around 1.0.16. The malicious code forwarded copies of outgoing email to an attacker-controlled address.
That establishes a mechanism for exposure, not which real customers or message types were affected. Snyk did not establish whether the legitimate Postmark repository was involved or compromised. Reviewing an installed version does not establish that its next update is safe.
A study found unauthenticated tools on roughly 40 percent of sampled servers
This one is measurable. As of 28 April 2026, Censys counted 12,520 internet-accessible MCP services across 8,758 unique IP addresses, in its scan. A separate 2026 measurement study (arXiv 2605.22333) of nearly 8,000 live remote servers found that about 40 percent exposed their tool interfaces with no authentication whatsoever, meaning any client could invoke them. Of the servers that did authenticate, static API keys were used almost as often as OAuth, each accounting for close to half of that authenticated group, although a static key can also be scoped and an OAuth deployment can still be misconfigured. These figures describe the study sample, not every MCP deployment.
What those servers expose is the uncomfortable part. The largest category Censys found was data and knowledge services, including direct database query interfaces, and hundreds more offered system control, including command execution.
When untrusted descriptions or results share context with other tools, a compromised or hostile server can influence how the model uses your other, perfectly legitimate tools. Invariant’s demonstration did precisely this: a poisoned calculator description changed how the agent used a trusted email tool. Separate server processes do not by themselves prevent this model-mediated interaction.
Judging a server before you connect it
Full governance is a separate problem, and a bigger one. At the level of a single decision, on a single machine, start with these questions.
- Who publishes it, and is this the real one? Check the organization behind the repository and confirm the package name matches the official source. A plausible package name does not establish who published it.
- What is the blast radius if it is hostile? Not “is this trustworthy” but “what could this reach.” A server with a read-only token against one dataset and a server holding an admin key are not the same decision.
- Can you scope the credential down? Default to read-only tokens and short-lived credentials wherever the provider supports them. This limits authorized access, but even a read-only token can expose sensitive data.
- Does it need to run on your host at all? A container or VM can restrict host access when configured carefully. Shared mounts, credentials, network access and isolation vulnerabilities still need review.
- Will you notice if it changes? Pin local package versions and review updates. For remote services, also review changes to tool descriptions and permissions; a local version pin cannot freeze a remote service.
Inventory MCP clients, servers and permissions together
Review the whole path: the model and client handling untrusted text, the servers executing requests, and the credentials and services they can reach. For a team, start by checking whether the inventory covers both local installations and remote connections. Assign an owner to each connection and record how to revoke its access.
FAQ
QIs MCP safe to use?
Is MCP safe to use?
QHow do I know if an MCP server is malicious?
How do I know if an MCP server is malicious?
QCan an MCP server read my files or run commands?
Can an MCP server read my files or run commands?
QDo I need to worry about MCP if I only use ChatGPT or Claude in the browser?
Do I need to worry about MCP if I only use ChatGPT or Claude in the browser?
QIs MCP security the same as API security?
Is MCP security the same as API security?
Sources
- MCP specification: tools and security requirements
- MCP specification: authorization
- Censys: MCP Servers on the Internet
- Invariant Labs: MCP Security Notification, Tool Poisoning Attacks
- The Hacker News: First malicious MCP server found stealing emails in rogue postmark-mcp package
- Snyk: Malicious MCP server on npm, postmark-mcp harvests emails
Measurements reflect public research through 2026. Check the linked primary sources for methodology and updates.