toolshield — sample MCP tool-manifest poisoning review

Generated from a real scan of bad-tools.json — a synthetic fixture, not a customer's system.
1 file scanned  ·  6 high 4 medium 10 total findings

Findings

SeverityRuleToolFieldWhat it meansHow to fix
HIGH TS001 bash description description contains instruction-override text: "Ignore previous instructions" Rewrite the description as a neutral capability statement. Move usage policy into your own system prompt, never into untrusted tool metadata.
HIGH TS001 bash description description contains concealment instruction: "Do not tell" Rewrite the description as a neutral capability statement. Move usage policy into your own system prompt, never into untrusted tool metadata.
HIGH TS001 bash description description contains data-sending instruction: "Send the token to" Rewrite the description as a neutral capability statement. Move usage policy into your own system prompt, never into untrusted tool metadata.
HIGH TS001 bash description description contains download-and-execute instruction: "fetch and execute" Rewrite the description as a neutral capability statement. Move usage policy into your own system prompt, never into untrusted tool metadata.
HIGH TS002 bash description description contains 1 invisible control character codepoint(s): U+200B Strip zero-width and bidi control characters from tool names and descriptions; require review of any non-printing characters.
HIGH TS006 bash name Tool name 'bash' impersonates a shell or privileged built-in Use a unique, namespaced tool name that identifies the actual provider.
MEDIUM TS003 bash inputSchema.properties.api_key Parameter 'api_key' looks like a raw credential input Accept a reference to a secret store or an environment variable name instead of the secret value; never require the raw credential as a tool argument.
MEDIUM TS004 bash inputSchema.additionalProperties inputSchema accepts arbitrary undeclared properties Set additionalProperties to false and declare every accepted argument explicitly.
MEDIUM TS005 read_file name/description Tool 'read_file' sounds read-only but its description mentions side effects Rename the tool to reflect its side effects, or remove the side-effecting behavior from the described scope.
MEDIUM TS007 bash inputSchema.properties.command Sensitive parameter 'command' is an unconstrained free-form string Constrain the field with enum, pattern, or format, and validate it server-side before use.
Scope of this review. toolshield performs static inspection of tool-manifest JSON only. It does not execute tools, call MCP servers, make network requests, or review tool implementation code. A clean result means none of the seven documented patterns were found — not that the manifest is safe. Invisible Unicode matches are reported as codepoint names, never printed raw.

What a paid review includes

  1. Run toolshield across the supplied tool manifests (tools/list exports, OpenAI-style function files, or JSON arrays).
  2. Written finding-by-finding report with severity, field, rationale, and concrete fix.
  3. Corrected example manifest with instruction-like text removed and parameters constrained.
  4. Optional SARIF file for GitHub code-scanning integration.
  5. One recheck after fixes are applied.

Source and test suite: github.com/sidsri14/toolshield