biopharma-patent-landscape

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

• Third-party content exposure detected (medium risk: 0.30). 在 Step 5/7/9 等路径中，工作流会对从专利检索返回的专利号/详情页面调用 patent.fetch 下载专利“md/全文”内容并让 LLM 读取与生成（如从 md 提取日期、靶点、以及基于 md 全文生成摘要），而这些文本虽然来自专利数据库但属于可被外部（第三方发布的专利公开文本）影响的免费文本。

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

• Potentially malicious external URL detected (high risk: 0.90). The skill clearly downloads patent MD full-text at runtime (via patent.fetch / F-column patentId links) using the Eureka URLs (e.g. https://eureka.zhihuiya.com/view/#/fullText?patentId={patent_id} and https://eureka.zhihuiya.com), and that fetched content is injected into the AI summarization pipeline, so external content fetched at runtime can directly influence prompts/results.