No prompt markup such as <IMPORTANT> or <system>
Tags such as <IMPORTANT>, <system> or <instructions>, and chat-template tokens such as [INST] or <|im_start|>, address the model directly and are the signature of published tool-poisoning attacks; describe the tool instead.
Check sec-instruction-markup · Poisoning and hidden text · default severity: blocking · rubric 2.1.0
What fails
Tags such as <IMPORTANT>, <system> or <instructions>, and chat-template tokens such as [INST] or <|im_start|>, are how published tool-poisoning attacks address the model directly.
The rule
Describe what the tool does, and don’t tell Claude how to behave.
How to fix it
Remove the markup and describe only what the tool does.
- Remove prompt-style tags and chat-template tokens from all metadata.
- Rewrite the content as a plain description of what the tool does.
- Add a CI test.
A CI test that fails when the pattern comes back, run against the saved tools/list answer:
TypeScript (node:test)
// tools-list.test.mjs: run `node --test` after saving the answer with
// npx @modelcontextprotocol/inspector --cli <server> --method tools/list > tools.json
import assert from 'node:assert/strict';
import { readFileSync } from 'node:fs';
import { test } from 'node:test';
const answer = JSON.parse(readFileSync('tools.json', 'utf8'));
const tools = answer.tools ?? answer.result?.tools ?? answer;
// Every string and every key in a tool, with its path.
function* strings(value, path) {
if (typeof value === 'string') yield [path, value];
else if (value && typeof value === 'object')
for (const [k, v] of Object.entries(value)) {
yield [`${path}.${k} (key)`, k];
yield* strings(v, `${path}.${k}`);
}
}
const PATTERN = /<\/?(important|system|instructions?|assistant)\b|\[\/?INST\]|<\|im_(start|end)\|>/i;
test('no prompt markup', () => {
const hits = [];
for (const tool of tools) for (const [path, text] of strings(tool, tool.name)) if (PATTERN.test(text)) hits.push(path);
assert.deepEqual(hits, [], 'prompt markup addresses the model instead of describing the tool');
});Python (pytest)
# test_tools_list.py: run `pytest` after saving the answer with
# npx @modelcontextprotocol/inspector --cli <server> --method tools/list > tools.json
import json
import re
with open("tools.json", encoding="utf-8") as f:
answer = json.load(f)
tools = answer.get("tools") or answer.get("result", {}).get("tools") if isinstance(answer, dict) else answer
def strings(value, path):
"""Every string and every key in a tool, with its path."""
if isinstance(value, str):
yield path, value
elif isinstance(value, dict):
for k, v in value.items():
yield f"{path}.{k} (key)", k
yield from strings(v, f"{path}.{k}")
elif isinstance(value, list):
for i, v in enumerate(value):
yield from strings(v, f"{path}.{i}")
PATTERN = re.compile(r"(?i)</?(important|system|instructions?|assistant)\b|\[/?INST\]|<\|im_(start|end)\|>")
def test_no_prompt_markup():
hits = [path for tool in tools for path, text in strings(tool, tool["name"]) if PATTERN.search(text)]
assert hits == [], "prompt markup addresses the model instead of describing the tool"Questions
We use <IMPORTANT> to stress a real constraint. Is that wrong?
Yes: state the constraint plainly ("Requires a project ID"). Anthropic asks descriptions to describe what the tool does, not tell Claude how to behave.
Where did this pattern come from?
Security researchers published tool-poisoning examples in 2025 that used <IMPORTANT> blocks to make models read files and leak them.
Related checks
- No wording that overrides instructions or hides actions from the user
- No HTML comments in metadata
- Metadata does not seek secrets or local files
- All checks in Tool poisoning and hidden text
How this check is graded
It counts toward: Poisoning and hidden text, Anthropic connector directory, OpenAI plugin directory. See the rubric.
Last verified against Anthropic connector pre-submission checklist 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