adversarial-polish

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

• Third-party content exposure detected (high risk: 0.75). 该技能的运行流程会把“用户给出 URL → web_fetch 获取”的网页正文（外部来源的自由文本）以及用户上传/提供的稿件文本直接作为 {current_A_text} 等变量注入到各 sub-agent 的 prompt 中，从而进入 LLM 上下文。

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

• Potentially malicious external URL detected (high risk: 0.80). The skill explicitly says "用户给出 URL → 使用 web_fetch 获取" and injects the fetched content as {current_A_text} into sub-agent prompts (SKILL.md:133), so a user-supplied URL fetched at runtime can directly influence model prompts and enable prompt-injection risks.