Oathe Security Badge

oathe / scan / reports / caution

Is wind-alice/alicemarket safe?

Use with caution. wind-alice/alicemarket is not clearly unsafe, but the audit flagged risks to review before you install it. Oathe's behavioral security audit gave the alicemarket skill by wind-alice a trust score of 70/100 with 8 findings (2 high). Report updated

https://github.com/wind-alice/alicemarket

70
CAUTION

Is wind-alice/alicemarket safe to install?

wind-alice/alicemarket is a large collection of professional financial analysis skills from Wind Information Co. Ltd (万得), a legitimate Chinese financial data company. The skills are well-structured and focused on legitimate equity research and trading analysis use cases, but four of the included skills contain silent auto-update mechanisms that download and execute code from an external GitHub repository after every invocation without user awareness or consent, creating a persistent supply chain attack surface. The wind-find-finance-skill additionally hard-instructs the AI agent to execute a Node.js update script before every finance query, constituting a mandatory pre-execution hook injected into the agent's reasoning loop.

What security issues were found in wind-alice/alicemarket?

Category Scores

Prompt Injection 65/100 · 30%
Data Exfiltration 70/100 · 25%
Code Execution 60/100 · 20%
Clone Behavior 90/100 · 10%
Canary Integrity 95/100 · 10%
Behavioral Reasoning 55/100 · 5%

Findings (8)

HIGH Silent auto-update installs code from GitHub without user consent -30 ▶

wind-one-mcp-skill, wind-alice, wind-mcp-skill, and wind-find-finance-skill all contain update-check.mjs scripts that automatically download and replace CLI code from https://github.com/Wind-Alice/AliceMarket.git after each successful skill invocation. The validation checks (syntax, version, transport type) are insufficient to detect semantically malicious updates that pass syntactic validation. The auto-update uses exponential backoff to persistently retry failed updates.

Evidence: wind-one-mcp-skill/SKILL.md: '本 Skill 内置自动升级。search、discover 或 call 返回正常 MCP 结果后,CLI 自动在后台检查并安装更新,无需用户执行更新命令或确认' and '更新固定使用主站仓库 https://github.com/Wind-Alice/AliceMarket.git'

HIGH Mandatory pre-execution hook forces agent to run Node.js before every finance query -25 ▶

wind-find-finance-skill SKILL.md contains a bolded mandatory instruction that the AI must execute an update-check.mjs script before reading any catalog or answering any finance question. This creates a persistent execution hook injected into the agent's reasoning loop on every finance-related interaction.

Evidence: wind-find-finance-skill/SKILL.md: '**必须先运行更新检查脚本,不得跳过。**在读取 catalog、判断 skill、回答用户或进入安装流程之前,必须按下面顺序查找更新脚本;找到第一个存在的路径后立即执行 node <path>'

MEDIUM Financial queries and API keys transmitted to external Wind servers -20 ▶

The wind-mcp-skill and wind-one-mcp-skill CLI scripts transmit user financial queries and API bearer tokens to https://mcp.wind.com.cn/vserver_wind_tools/mcp/ on every invocation. The wind-alice skill transmits raw user prompts to https://api.windalice.com. While this is the stated purpose, it creates exfiltration channels that would be trivially weaponized if CLI scripts are modified via the auto-update mechanism.

Evidence: wind-one-mcp-skill/SKILL.md: 'CLI 连接 https://mcp.wind.com.cn/vserver_wind_tools/mcp/,处理 Bearer 认证'; wind-alice/SKILL.md: '禁止改写用户问句:把用户的原话作为 --prompt 传入即可'

MEDIUM Skills instruct agent to install additional software from external repositories -15 ▶

Multiple skills contain explicit instructions directing the AI agent to execute 'npx skills add Wind-Information-Co-Ltd/wind-skills ...' or Gitee mirror equivalents to install other skills from external sources. The wind-find-finance-skill enforces this for missing skills and prohibits the agent from answering until installation completes.

Evidence: wind-find-finance-skill/SKILL.md: '在用户确认或拒绝前,不得...由 AI 直接安装,不要求用户复制或手动执行命令'; multiple SKILL.md files contain npx skills add commands pointing to github.com and gitee.com

MEDIUM Four independent auto-update scripts from same GitHub source create broad attack surface -15 ▶

The repository contains four separate update-check.mjs scripts across wind-one-mcp-skill, wind-alice, wind-mcp-skill, and wind-find-finance-skill. All target the same GitHub repository. A single compromise of the Wind-Alice/AliceMarket GitHub repository would propagate to all four skill components simultaneously.

Evidence: Files: skills/wind-alice/scripts/update-check.mjs, skills/wind-one-mcp-skill/scripts/update-check.mjs, skills/wind-mcp-skill/scripts/update-check.mjs, skills/wind-find-finance-skill/scripts/update-check.mjs

LOW Skills adopt strong expert personas that override default AI identity -5 ▶

Several skills define System Prompt sections instructing the AI to fully adopt specific expert personas: '你是一名中国A股顶级市场结构研究员,拥有15年以上市场交易与盘面分析经验' and avatar skills for Buffett, Munger, Taleb, Naval Ravikant. While common for domain-specific skills, strong persona adoption can subtly bias financial advice.

Evidence: a-share-primary-theme-identification/SKILL.md System Prompt section; avatar-warren-buffett-investing/SKILL.md; avatar-nassim-taleb-risk/SKILL.md; avatar-charlie-munger-thinking/SKILL.md; avatar-naval-ravikant-thinking/SKILL.md

INFO Sensitive file paths accessed before skill clone but canary check clean -5 ▶

Inotifywait recorded reads of .env, .ssh/id_rsa, .aws/credentials, .npmrc, .docker/config.json, and GCP credentials at 07:44:32, approximately 5 seconds before the GitHub clone began at 07:44:37. Timeline and process context (sudo, pid 1103) indicate these reads were from the audit framework establishing its monitoring baseline, not from the skill. All canary files confirmed intact.

Evidence: inotifywait: '07:44:32 OPEN /home/oc-exec/.env'; auditd PATH records at audit(1794815072.128-135); 'Monitoring: Canary File Integrity: ✅ All canary files intact'

INFO Clone behavior normal — no unexpected network connections or process spawning -5 ▶

The git clone connected only to GitHub (140.82.121.3:443) via HTTPS and used expected git sub-processes (git-remote-https, index-pack, rev-list). No unexpected destinations were contacted. No network state changes were observed in the before/after comparison beyond the SSH session change.

Evidence: Network log: 'ens3 Out IP 10.0.2.15.46036 > 140.82.121.3.443'; EXECVE: 'git clone --depth 1 -- https://github.com/wind-alice/alicemarket'; Connection diff: only SSH port changed

Should I install wind-alice/alicemarket?

Oathe's verdict for wind-alice/alicemarket is CAUTION with a trust score of 70/100. Recommendation: Review before install.