Research Ideation

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

• Third-party content exposure detected (high risk: 0.80). The skill's required workflow explicitly performs WebSearch across public academic sites (arXiv, Google Scholar) and auto-imports/attaches open-access PDFs via the Zotero MCP flow (add_items_by_doi, find_and_attach_pdfs, get_item_fulltext), meaning the agent ingests and reads untrusted third‑party web content as part of its analysis.

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

• Potentially malicious external URL detected (high risk: 1.00). The skill invokes Zotero MCP and WebSearch at runtime to fetch paper content (e.g., DOI/arXiv links such as https://doi.org/10.48550/arXiv.xxxx.xxxxx and https://arxiv.org/abs/xxxx.xxxxx) and then calls get_item_fulltext to read and inject full-text into the agent's analysis/context, so external URLs are fetched at runtime and directly influence prompts/outputs which the skill depends on.