What is MCP security?
MCP security is the practice of controlling the risks that Model Context Protocol servers and clients add to an environment: servers that answer without authentication, credentials held by servers, tool descriptions that carry instructions, malicious or compromised server packages, and tool calls that reach more data than the task needs.
Also called Model Context Protocol security, MCP server security. Last reviewed .
Why it matters
The tooling around MCP has had critical vulnerabilities. CVE-2025-6514 is an operating system command injection in mcp-remote, triggered by connecting to an untrusted MCP server. CVE-2025-49596 allowed remote code execution in MCP Inspector versions below 0.14.1, because the Inspector client and its proxy did not authenticate each other.
Server packages have been impersonated. On 25 September 2025, Postmark warned that an npm package named postmark-mcp, which Postmark did not publish, had added a hidden BCC in version 1.0.16 that copied every email it sent to an outside address.
Many servers are simply open. Guard0's July 2026 census reached 2,958 remote MCP servers listed in the public MCP Registry. 1,139 of them (39%) answered with no authentication. The specification's own security best practices cover the confused deputy problem, token passthrough, session hijacking and compromise of local servers.
Common risks
- Servers that answer without authentication
- The specification says "Authorization is OPTIONAL for MCP implementations", so a remote server can be published with none. Guard0's July 2026 census reached 2,958 remote MCP servers listed in the public MCP Registry. 1,139 of them (39%) answered with no authentication.
- Tool poisoning
- A tool description carries instructions written for the model. The OWASP MCP Top 10 lists it as MCP03:2025, and tool poisoning has its own entry here.
- Impersonated or compromised server packages
- A server is code someone installs. Postmark warned in September 2025 that the npm package postmark-mcp, which it did not publish, copied every email it sent to an outside address from version 1.0.16.
- Vulnerable clients and tooling
- CVE-2025-6514 is a command injection in mcp-remote, triggered by connecting to an untrusted server. CVE-2025-49596 allowed remote code execution in MCP Inspector versions below 0.14.1.
- Confused deputy and token passthrough
- A proxy server uses its own authority for the wrong client, or forwards a token it was never issued. The security best practices cover both, and confused deputy has its own entry.
- Servers nobody approved, and calls nobody recorded
- The OWASP MCP Top 10 lists MCP09:2025 Shadow MCP Servers and MCP08:2025 Lack of Audit and Telemetry. An unapproved server is one form of shadow AI.
Related terms
- Model Context Protocol (MCP)
- The Model Context Protocol is the protocol being secured.
- Tool poisoning
- Tool poisoning is the MCP attack that works through tool descriptions.
- Confused deputy
- A confused deputy arises when an MCP proxy uses its own authority for the wrong client.
- Shadow AI
- An MCP server nobody approved is one form of shadow AI.
- AI bill of materials (AIBOM)
- An AI bill of materials records which MCP server packages and versions are in use.
How Guard0 handles it
Guard0 lists every MCP server your agents connect to, with the tools each one exposes and the credential it holds: a static key, OAuth, or none. MCP declarations in your repositories are checked with each pull request. Where you turn guards on, a guard can allow, redact, coach or deny a tool call, and every decision goes on the record.
Read the guide: MCP Security Guide: Risks, Controls and a Checklist
Questions
Is MCP secure by default?
No. The protocol leaves the main controls to whoever deploys it. The specification states that "Authorization is OPTIONAL for MCP implementations", and the tools section says there "SHOULD always be a human in the loop with the ability to deny tool invocations". Whether a server asks for credentials, and whether a person approves a call, depends on each server and each client.
How is MCP security different from API security?
API security protects an endpoint from the callers outside it. With MCP the caller is an AI agent that chooses its tools by reading their descriptions, so a changed description can redirect it. MCP security also checks the server, its tools, the credential it holds and each call the agent makes.
How many MCP servers are exposed without authentication?
Guard0's July 2026 census reached 2,958 remote MCP servers listed in the public MCP Registry. 1,139 of them (39%) answered with no authentication. The report gives the method and the full figures.
Which MCP vulnerabilities have a CVE?
Three that are cited in this glossary. CVE-2025-6514 is an operating system command injection in mcp-remote. CVE-2025-49596 is remote code execution in MCP Inspector below version 0.14.1. CVE-2025-54136 let an attacker replace an approved MCP configuration in Cursor 1.2.4 and below without a new prompt.
Sources
- MCP specification: Security Best Practices. Model Context Protocol.
- MCP specification: Authorization. Model Context Protocol.
- MCP specification: Tools. Model Context Protocol.
- CVE-2025-6514. NVD.
- CVE-2025-49596. NVD.
- CVE-2025-54136. NVD.
- Information regarding malicious postmark-mcp package. Postmark.
- OWASP MCP Top 10. OWASP.
- The State of Exposed Agents, July 2026. Guard0 Research.