hv-analysis

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

• Third-party content exposure detected (high risk: 1.00). The SKILL.md explicitly requires "必须联网搜索" and instructs child agents to use WebSearch/WebFetch and external sources (e.g., GitHub issues, Reddit, Twitter/X, 知乎, arXiv API via curl) to fetch and ingest public/user-generated webpages as core inputs to the analysis, so untrusted third‑party content will be read and used to drive decisions in the workflow.

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

• Potentially malicious external URL detected (high risk: 1.00). The skill's runtime sub-agent prompts explicitly instruct fetching arXiv data via the external API (e.g. https://export.arxiv.org/api/query?search_query=...) and subsequent arxiv.org pages (https://arxiv.org/abs/...), which would inject externally fetched content into the agent's context and is required for academic research, so it directly controls prompts at runtime.