MCP Tool Poisoning (Model-Directed Instructions in Tool Metadata)
mcp-tool-poisoning is a high severity check in the
ai category, written in python.
Its source is templates/ai/mcp/mcp-tool-poisoning.py in
cert-x-gen-templates.
What it checks
Section titled “What it checks”Detects MCP servers whose tool, prompt or resource descriptions embed instructions directed at the connected model rather than at the user - instructions enclosed in a pseudo-tag or an HTML comment, imperative chains, and text concealed in invisible Unicode
Header
Section titled “Header”| Field | Value |
|---|---|
| Id | mcp-tool-poisoning |
| Severity | high |
| Language | python |
| Category | ai |
| Author | Bugb Research |
| Template version | 2.0.0 |
| Confidence | 90 |
| CVSS | 8.2 |
| Weakness | CWE-1427 |
| Tags | mcp, ai, agent, tool-poisoning, prompt-injection, cwe-1427 |
| Target kind | http |
| Oracle | property |
References
Section titled “References”Declared in the header. cxg parses @references and then discards it, and
nothing at scan time reads it, so this is the only place the links a
template cites are surfaced.
- https://modelcontextprotocol.io/specification
- https://cwe.mitre.org/data/definitions/1427.html
- https://trojansource.codes/
Run it
Section titled “Run it”To see what the copy on your machine says about itself, and to confirm it is installed at all:
cxg template info mcp-tool-poisoningThe id it prints is the one to pass anywhere a template is selected. See
cxg template for the rest of
the subcommand, Scan a target for running
a scan, and
A match is not a finding for how
to read what comes back.
Related
Section titled “Related”- Template catalog is the whole corpus.
- ai is the rest of this category.
- Templates by weakness class groups checks by the defect they look for.
- Contribute a template is the route from a check you needed and could not find to a merged one.