zg-wewrite

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

• Third-party content exposure detected (high risk: 0.78). Step 3.2 “素材采集 + 内容增强”会进行运行时 WebSearch/抓取外部网页内容（如 mp.weixin.qq.com、36kr、v2ex、zhihu 等），这些网页的正文/引述会被提取并作为 Step 4 写作的输入进入 LLM 上下文，属于“公共网页内容/外部文章文本”间接提示注入风险。

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

• Potentially malicious external URL detected (high risk: 0.90). The skill explicitly runs scripts that fetch a user-supplied article URL at runtime (e.g., python3 .../scripts/fetch_article.py && python3 .../scripts/extract_exemplar.py ...) and then injects the fetched article text as exemplars into the writing prompt (see SKILL.md writing-persona/exemplar injection), so external web content provided via that is fetched at runtime and directly controls model prompts.