study

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

• Third-party content exposure detected (high risk: 0.90). The skill explicitly downloads and parses PDFs from arbitrary HTTP(S) URLs (see SKILL.md Step 1a and scripts/download-pdf.cjs), then feeds the extracted text into parse-pdf.js and downstream steps (Step 2/Step 3 tag generation, assessments, and file/code generation), so untrusted third‑party PDF content can directly influence decisions and actions.

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

• Potentially malicious external URL detected (high risk: 0.80). The skill downloads arbitrary PDFs at runtime (example: https://arxiv.org/pdf/1706.03762.pdf) and then parses and injects their text into the agent's processing pipeline, so externally-hosted document content can directly influence the model's context and outputs.