The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted scholarly metadata and text (such as abstracts, author names, and paper content) from multiple external APIs (PubMed, CrossRef, ArXiv, Scopus, ScienceDirect) and user-provided manuscripts. This creates a surface for indirect prompt injection where malicious instructions embedded in scholarly records could attempt to influence the agent's behavior.
Ingestion points: API responses processed in mcp-server/sources/ and document extraction in references/workflows/wf2-citation-verification.md.
Boundary markers: No specific delimiters or instructions are provided to help the agent distinguish between its instructions and the potentially untrusted scholarly data it retrieves.
Capability inventory: The skill possesses network access via the MCP server and file-writing capabilities (e.g., exporting citation files like .ris and .bib via scripts/format-converter.py).
Sanitization: The skill uses defusedxml to prevent XML-based attacks and employs ris_escape in scripts/converters.py to strip HTML tags from citation metadata.
[DYNAMIC_EXECUTION]: The installation script employs inline Python commands to programmatically update the agent's local configuration files during setup.
Evidence: Found in install.sh where python3 -c is used to merge server settings into the agent's .mcp.json and settings.json files.
Context: This is a routine administrative task used during the installation of AI agent skills to ensure the MCP server is correctly registered.
[DATA_EXFILTRATION]: The skill is configured to access the local pybliometrics configuration file (~/.config/pybliometrics.cfg) to retrieve Elsevier/Scopus API credentials for authorized scholarly searches.
Evidence: This access is documented in README.md and implemented in mcp-server/sources/elsevier_common.py.
Context: This represents the standard, documented method for the pybliometrics library to manage credentials locally. The skill explicitly advises users to manage their keys via this local configuration rather than hardcoding them in the repository.

nature-academic-search