Skip to content
IsoKron

No steering of when other tools are used

A tool description must not steer the model toward or away from other servers’ tools ("always call this tool first", "never use other tools"); OpenAI and Anthropic both reject it, while routing between your own tools is fine.

Check sec-cross-tool-steering · Poisoning and hidden text · default severity: attention (or blocking, depending on the case) · rubric 2.1.0

What fails

Wording such as "always call this tool first", "before using any other tool" or "never use other tools" interferes with other tools and is blocking; Anthropic also rejects descriptions that "interfere with Claude calling other tools". Softer wording such as "use this instead of X" is marked for attention. Routing between tools of the same server ("call get_issue first", "prefer this tool over search_issues" when search_issues is in the same list) and ordering inside the server ("call this first to get the instance name") are fine.

The rule

Plugins must not manipulate how the model selects or uses other plugins or their tools (for example, instructing the model to prefer one plugin over others), interfere with fair discovery, or otherwise negatively impact ChatGPT or our users.

Source: OpenAI plugin guidelines, section 3.4

How to fix it

Say what this tool is for and what it is not for, without directing the model away from other tools.

  1. Remove wording that ranks this tool above others or forbids other tools.
  2. Keep routing within your own server ("call get_issue first for the version").
  3. Say what this tool is for and not for, without naming tools you do not own.

TypeScript SDK

TypeScript
import { McpServer } from '@modelcontextprotocol/server';
import * as z from 'zod';

const server = new McpServer({ name: 'issues', version: '1.0.0' });

server.registerTool(
  'search_issues',
  {
    title: 'Search issues',
    description: 'Search issues by text. Use when the user describes an issue but has no ID; for a known ID call get_issue.',
    inputSchema: z.strictObject({
      query: z.string().describe('Words to search for'),
    }),
    annotations: { readOnlyHint: true, destructiveHint: false, openWorldHint: false },
  },
  async ({ query }) => ({ content: [{ type: 'text', text: 'Done' }] }),
);

Python SDK

Python
from typing import Annotated

from pydantic import Field
from mcp.server.mcpserver import MCPServer
from mcp.types import ToolAnnotations

server = MCPServer("issues")

@server.tool(
    title="Search issues",
    annotations=ToolAnnotations(readOnlyHint=True, destructiveHint=False, openWorldHint=False),
)
def search_issues(
    query: Annotated[str, Field(description="Words to search for")],
) -> str:
    """Search issues by text. Use when the user describes an issue but has no ID; for a known ID call get_issue."""
    return "Done"

Run against @modelcontextprotocol/server 2.4.0, mcp 2.3.0 and fastmcp 4.2.2 on 2026-11-18; the checker confirms this check passes on the tool list each snippet produces.

Questions

What is tool shadowing?

A description that changes how the model uses other tools, for example telling it to add a hidden recipient whenever any email tool is called. The description of one server reaches into another.

Can I say "prefer this over X"?

Only if X is in your own server. Steering away from another server’s tools is what both directories reject.

How this check is graded

It counts toward: Poisoning and hidden text, Anthropic connector directory, OpenAI plugin directory. See the rubric.

Last verified against OpenAI plugin guidelines, section 3.4 on 2026-11-18. The rule above is quoted in the source’s own words. This page as Markdown.

Check your tool list: paste your tools/list, nothing is sent anywhere