dev.latamref/reference-data

What we checked

Six static checks, weighted by risk. Every result reflects only what could be observed in the published package and repository — never intent.

• Injection surface

  warn25% of grade

  Tool descriptions/manifest scanned for instruction-injection patterns (imperatives at the model, hidden text, 'ignore previous', data-exfil URLs).

  tool "compute_income_tax" description: bare 'IMPORTANT:' prose in a description (tag form is high-confidence; prose is warn).

• Supply chain

  warn25% of grade

  Package provenance: namespace verification, repo linkage, maintainer count, account age, postinstall scripts, typosquat distance.

  no resolvable repository linked to the package (provenance unverifiable)

• Credential hygiene

  fail15% of grade

  How the server takes secrets (env vs plaintext config vs hardcoded); secrets appearing in tool schemas.

  tool "get_series" has a secret-shaped input parameter "api_key" — secrets must never be tool parameters (the model would supply/handle the credential). tool "get_series_history" has a secret-shaped input parameter "api_key" — secrets must never be tool parameters (the model would supply/handle the credential). tool "check_minimum_wage" has a secret-shaped input parameter "api_key" — secrets must never be tool parameters (the model would supply/handle the credential). tool "compute_vat" has a secret-shaped input parameter "api_key" — secrets must never be tool parameters (the model would supply/handle the credential). tool "count_working_days" has a secret-shaped input parameter "api_key" — secrets must never be tool parameters (the model would supply/handle the credential). tool "compute_income_tax" has a secret-shaped input parameter "api_key" — secrets must never be tool parameters (the model would supply/handle the credential). tool "get_snapshot" has a secret-shaped input parameter "api_key" — secrets must never be tool parameters (the model would supply/handle the credential).

• Permission scope

  unscannable15% of grade

  Declared tools vs. breadth (filesystem, network, exec); flags shell-exec and unbounded filesystem access.

  No source files, package.json, or pyproject were fetched; cannot infer capability scope.

• Version behavior

  pass10% of grade

  Diff of tool definitions between versions; new permissions or changed descriptions in a patch release (the postmark-mcp class).

  tool surface, capabilities, install scripts, and publisher unchanged vs prior version 1.1.0 (minor bump); no rug-pull drift.

• Transport config

  fail10% of grade

  Remote servers: TLS and auth mode (none/token/OAuth). Local servers: whether the manifest indicates it phones home.

  remote MCP endpoint declares NO authentication (unauthenticated remote — nginx-ui/CVE-2026-33032 class)