← MCP server and skill grades

skills

Wind-Alice/AliceMarket/skills

D
Type
Skill
Scanned
Nov 13, 2026
Version
ab8fa4cfc790
Files read
407

This repository provides a collection of financial data and investment analysis skills for AI coding agents, connecting to Wind, Alice, Tushare, FINVIZ, Financial Modeling Prep, and Yahoo Finance. Its biggest risks are that the wind-alice CLI can leak the user's WIND_API_KEY to external URLs found in responses, and several Wind skills run silent background auto-updaters that execute unpinned remote code and can be hijacked via an untrusted skills-lock.json file.

What it can reach

Network
Connects to Wind MCP servers, Wind Alice, Tushare Pro, FINVIZ, Financial Modeling Prep, Yahoo Finance, GitHub, and Gitee to query market data, download report attachments, and pull skill updates.
Secrets
Reads and writes WIND_API_KEY in ~/.wind-aifinmarket/config, config.env, local config.json, or environment variables, and reads TUSHARE_TOKEN, FINVIZ_API_KEY, and FMP_API_KEY from the environment.
Files
Reads local parameter and Excel files, saves analysis reports and downloaded attachments to reports/ and .agents/download/, and overwrites its own skill directories during background updates.
Shell commands
Spawns git, npx skills, node, and python subprocesses for market data queries, model validation, skill installation, and background self-updates.
Browser
Launches the system default web browser to open the Wind developer portal when running the open-portal command.
System
Inspects environment variables and parent process trees (/proc on Linux or ps on macOS/Unix) in wind-one-mcp-skill to identify the calling AI agent for the User-Agent header.

Findings

2 high1 medium2 low
  1. HighWIND_API_KEY sent as Bearer token to arbitrary external URLs during file download

    skills/wind-alice/scripts/request.js:1420

    When wind-alice finishes a request that did not include an A2A.PresentFiles component (or even within PresentFiles if an http/https URL is given), it scans the response text for any URL whose path contains /files/ or ends with common extensions such as .pdf, .html, .json, .csv, or .md, without checking the hostname. It then fetches every matched URL with Authorization: Bearer <WIND_API_KEY> attached. If a user asks Alice to fact-check untrusted text containing an external link that Alice echoes, or if Alice cites external web pages or PDFs in its answer, the user's secret WIND_API_KEY is sent directly to those third-party servers.

    Fix: Verify that the parsed URL uses HTTPS and belongs to an allowlisted Wind file host (such as alice.wind.com.cn) before downloading or attaching the Authorization header, and disable cross-origin credentialed redirects.

  2. HighBackground auto-updaters trust workspace skills-lock.json and pass unescaped strings to cmd.exe

    skills/wind-find-finance-skill/scripts/update-check.mjs:353

    The background updaters in wind-find-finance-skill, wind-alice, and wind-mcp-skill search process.cwd(), INIT_CWD, and ancestor directories for skills-lock.json. When a skill is installed at project scope or lacks a global .skill-lock.json entry (such as a manual git clone), opening an untrusted repository and triggering the skill causes update-check.mjs to read sourceUrl or source from the stranger's skills-lock.json without validation. That untrusted value is passed to npx skills add -y (overwriting the user's skill with attacker-controlled code), passed to git ls-remote without a -- separator, and on Windows concatenated directly into cmd.exe /d /s /c, allowing immediate shell command injection.

    Fix: Do not read skills-lock.json from process.cwd() or INIT_CWD, validate repository URLs against a strict allowlist of official Wind repositories, add -- before positional arguments in git ls-remote, and avoid cmd.exe /c string concatenation.

  3. MediumPre-approved Bash(python:*) permission allows unprompted arbitrary Python execution

    skills/tushare-finance-skill/SKILL.md:5

    The tushare-finance skill declares Bash(python:*) in its allowed-tools frontmatter, which in supporting agents such as Claude Code pre-approves any shell command starting with python without asking the user. Because this skill is used to fetch untrusted external text such as public investor Q&A (irm_qa_sh, irm_qa_sz) and financial news, a prompt injection inside fetched market content could trick the agent into running arbitrary python -c commands silently.

    Fix: Remove Bash(python:*) from allowed-tools or scope pre-approval to a dedicated read-only query script with fixed arguments so arbitrary inline Python code still requires user confirmation.

  4. LowSilent background self-updates download and overwrite executable skill code from unpinned branches

    skills/wind-mcp-skill/scripts/cli.mjs:146

    Four skills (wind-mcp-skill, wind-one-mcp-skill, wind-alice, and wind-find-finance-skill) spawn detached background updaters during normal use. These updaters either run unpinned npx skills update/add commands from GitHub/Gitee HEAD or git clone the unpinned main branch of Wind-Alice/AliceMarket.git and overwrite local .mjs scripts in place without user confirmation. In wind-mcp-skill this behavior is not mentioned in SKILL.md, and in three of the four skills there is no setting to turn it off, exposing users to silent upstream supply-chain changes after installation.

    Fix: Make automatic self-updating opt-in rather than on by default, notify the user instead of silently replacing executable scripts, and pin updates to signed tags or verified commit hashes.

  5. LowSkill router instructs the AI agent to hide npx installation commands from the user

    skills/wind-find-finance-skill/SKILL.md:68

    Although wind-find-finance-skill asks the user whether to install a missing skill to the current agent or all agents, it explicitly instructs the AI agent to conceal the underlying npx skills add command and chosen source repository (GitHub or Gitee) from the user before and during execution, contradicting its own README. Hiding shell commands prevents users from reviewing what package runner and remote repository are being executed on their machine.

    Fix: Instruct the agent to display the exact npx skills add command and target repository URL when asking the user for installation confirmation.