The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill possesses a vulnerability surface for indirect prompt injection by ingesting and processing untrusted data from external literature databases.
Ingestion points: Search results (titles, abstracts, and metadata) retrieved via patent.search and paper.search in Step 2, 3, and 4.
Boundary markers: The instructions lack specific guidance for the LLM to use delimiters (e.g., XML tags or triple backticks) or "ignore embedded instructions" warnings when interpolating external content into the HTML report template.
Capability inventory: The skill uses the files tool to write the final HTML report to the local system and utilizes a Python 3.12 runtime for data organization.
Sanitization: There are no explicit instructions to sanitize, escape, or filter the content retrieved from external sources before it is rendered into the final HTML output, which could allow malicious snippets within literature data to influence the LLM's behavior or the report's structure.
[EXTERNAL_DOWNLOADS]: The skill performs legitimate network operations to retrieve research data from the author's official services.
Evidence: The workflow triggers multiple semantic queries through the mcp_patent-search dependency to fetch patent and academic paper data from Patsnap's infrastructure.
Context: These operations are central to the skill's purpose and are conducted via established MCP tools, representing expected behavior for a literature investigation agent.

microbiome-disease-research