skill-evo

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.90). The skill's research workflow (templates/command-research.md Step 2 "并行信息采集" and Step 3 "源码深度分析") explicitly instructs the agent to WebFetch official docs, GitHub READMEs, community sources (reddit/hackernews), npm/bundlephobia APIs and to clone arbitrary repos, meaning it fetches and ingests untrusted public web content that can influence recommendations and actions.

MEDIUM W013: Attempt to modify system services in skill instructions.

• Attempt to modify system services in skill instructions detected (high risk: 0.70). This skill explicitly instructs the agent to run an installation script that creates directories, copies hook scripts, updates .claude/settings.json and .gitignore, and registers persistent hooks that capture session data—i.e. it actively modifies project/global configuration and installs executable hooks on the machine (though it does not request sudo or create users), so it pushes the agent to change the machine state and poses a moderate-high risk.