legal-research

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

• Third-party content exposure detected (high risk: 0.75). SKILL.md 規定在步驟二/三會呼叫外部 MCP 工具（如 dr-lawbot:search_bundle 與 taiwan-legal-db:*）並把其回傳的判決內容/候選（含 allowed_citations、unread_candidates、case_history 等）作為 LLM 上下文用於引用驗證，因此屬於「運行時從外部來源讀取的自由文本/內容」類別的間接提示注入風險。

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.80). The skill instructs the agent to register and call a remote MCP at https://tlr.dr-lawbot.com/mcp at runtime (via "claude mcp add ... https://tlr.dr-lawbot.com/mcp" and the mcpServers entry), so responses and citation/trust decisions can be directly controlled by content returned from that URL.