legal-graph

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

• Third-party content exposure detected (medium risk: 0.30). 在 SKILL.md 步驟一/二會消費上游技能（特別是 legal-research 的 dr-lawbot search_bundle 回傳之 citation_text、cited_articles 等結構化文字與判決要旨），而此內容是在運行時由遠端工具讀取/返回；若攻擊者可影響該遠端來源所回傳的文字，便屬外部人可投毒的自由文本暴露。

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

• Potentially malicious external URL detected (high risk: 0.90). The skill includes explicit runtime registration and probe commands that call the remote MCP endpoint https://tlr.dr-lawbot.com/mcp (e.g., claude mcp add ... https://tlr.dr-lawbot.com/mcp, npx mcp-remote https://tlr.dr-lawbot.com/mcp, and a curl probe), which the agent would contact at runtime to fetch search_bundle results that directly steer its retrievals and prompt behavior.