MCP Server Security

MCP tool descriptions and schemas risks

Six risks sit in what the model reads: poisoned descriptions, shadowing across servers, hostile schemas and return values, trusted annotations, prompt injection in tool results and server-supplied schemas that burden clients. Each row cites OWASP, the MCP spec or published research, with severity from one written rule, read on 16 Nov 2026.

Facts read on .

What this dimension covers

What the model reads: descriptions, schemas, annotations and tool results. The register holds 6 rows here (R-23, R-24, R-25, R-26, R-27, R-28). Each row states what the risk is, its impact, concrete controls and at least one primary source.

Sources: OWASP MCP Security Cheat Sheet; When an MCP Tool Asks for Your GitHub Token, Knostic; AI tool poisoning through MCP tool descriptions, CrowdStrike.

Highest-severity rows

Rows that meet the high rule under the written rule are: R-23 Poisoned tool description; R-27 Prompt injection through tool results.

Sources: OWASP MCP Security Cheat Sheet; When an MCP Tool Asks for Your GitHub Token, Knostic; AI tool poisoning through MCP tool descriptions, CrowdStrike; MCP Tools, 2026-07-28 revision.

First controls to apply

Start with the controls on the high rows, then work down. Controls name settings only where a primary source documents them; check each against the version your clients run.

Sources: OWASP MCP Security Cheat Sheet; When an MCP Tool Asks for Your GitHub Token, Knostic; AI tool poisoning through MCP tool descriptions, CrowdStrike; MCP Tools, 2026-07-28 revision.

Tool descriptions and schemas: 6 items

R-23: Poisoned tool description

Instructions hidden in a tool description steer the model, which reads the description to decide how to call the tool. Knostic described a package on 12 Nov 2026 whose description tells the agent to fetch a GitHub token without asking the user.

Impact: The agent takes an action the user did not request, such as reading a credential.

Controls:

  • Review the full text of every tool description and parameter field before a server is approved.
  • Hash and pin the reviewed definitions and alert on change.
  • Scan descriptions with a tool built for the purpose (the OWASP cheat sheet names mcp-scan) and keep a human approval step for sensitive calls.

Severity: high

Sources: OWASP MCP Security Cheat Sheet (read ); When an MCP Tool Asks for Your GitHub Token, Knostic (read ); AI tool poisoning through MCP tool descriptions, CrowdStrike (read ).

R-24: Tool shadowing across servers

One server's tool text manipulates how the agent uses tools from another, trusted server, because the model sees all connected descriptions at once.

Impact: A low-trust server can alter how a high-trust one is used, including which data is sent where.

Controls:

  • Treat each server as its own security domain and keep sensitive servers apart from general-purpose ones.
  • Monitor for cross-server data flows, such as one server's credentials appearing in another's calls.
  • Limit how many servers a single agent session may load.

Severity: medium

Sources: OWASP MCP Security Cheat Sheet (read ).

R-25: Hostile content in schemas and return values

Instructions can sit in parameter names, types, defaults and the values a tool returns, not only in the description field.

Impact: Review that checks only the description misses the other fields the model reads.

Controls:

  • Inspect every field of the definition, including inputSchema and outputSchema.
  • Set additionalProperties to false and use pattern constraints on string fields.
  • Validate and sanitize tool outputs before they re-enter model context.

Severity: medium

Sources: OWASP MCP Security Cheat Sheet (read ); MCP Tools, 2026-07-28 revision (read ).

R-26: Annotations trusted as enforcement

A client lets a server's readOnlyHint or destructiveHint decide how much approval to ask for. The spec says clients must consider annotations untrusted unless they come from trusted servers.

Impact: A flipped hint can change approval behavior with no visible change to the description.

Controls:

  • Treat annotations as hints only and decide approval from your own policy per tool.
  • Include annotations in the pinned hash of each tool definition.
  • Require confirmation for sensitive reads, writes and irreversible actions whatever the label says.

Severity: medium

Sources: MCP Tools, 2026-07-28 revision (read ); OWASP MCP Security Cheat Sheet (read ).

R-27: Prompt injection through tool results

A tool returns text, such as a web page or ticket, that contains instructions the model then follows.

Impact: An approved server can still cause unapproved actions through the content it returns.

Controls:

  • Treat every tool response as untrusted data, including responses from approved servers.
  • Enforce authorization and validate follow-up tool calls in application code, regardless of the model's interpretation.
  • Extract structured fields from retrieval tools instead of passing raw pages.

Severity: high

Sources: OWASP MCP Security Cheat Sheet (read ).

R-28: Server-supplied schemas that exhaust or confuse clients

A client evaluates patterns or fetches references in a schema a server supplies. Inspector 2.14.0 now checks schema patterns in bounded time, and Python SDK advisory GHSA-rwrf-2pqf-9j8j describes a client fetching server-chosen $ref URLs.

Impact: A hostile server can slow or break a client, or cause it to make requests it should not.

Controls:

  • Evaluate server-supplied patterns with a linear-time matcher and limits.
  • Do not resolve remote references from schemas supplied by untrusted servers.
  • Keep MCP clients and the Inspector on current releases.

Severity: low

Sources: MCP Inspector 2.14.0 release notes (read ); GHSA-rwrf-2pqf-9j8j: MCP Python SDK server-chosen $ref URLs (read ).