paper-deep-reading

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

• Third-party content exposure detected (medium risk: 0.30). In scripts/fetch.py → resolve_europepmc() / resolve_preprint() the workflow fetches outsider-authored paper metadata/abstract and then downloads and parses the PDF bytes from user-supplied arXiv/DOI/PMCID/PMID URLs, making the paper text (including any adversarial content inside the PDF) directly readable by the runtime LLM steps that “teach/judge/write” the note.