The Agent Skills Directory

EXTERNAL_DOWNLOADS (MEDIUM): The skill instructs users to install arxiv-mcp-server from an untrusted source (blazickjp/arxiv-mcp-server) using uv or npx. This package is not from a trusted GitHub organization or repository as defined in the security policy.
COMMAND_EXECUTION (LOW): The documented configuration for the ArXiv MCP server involves executing code via uv tool run, which runs the unverifiable package mentioned above.
DATA_EXFILTRATION (LOW): The skill facilitates network requests to api.exa.ai for search and synthesis. While consistent with the skill's purpose, this domain is not on the trusted whitelist for network operations.
PROMPT_INJECTION (LOW): The skill is vulnerable to Indirect Prompt Injection (Category 8). It ingests untrusted data from external sources (academic papers via read_paper and web results via Exa) and processes it with a deep analysis prompt without explicit boundary markers or sanitization.
Ingestion points: mcp__exa__web_search_exa, search_papers, and read_paper (which extracts full text from arXiv PDFs).
Boundary markers: Absent; the skill does not wrap external content in delimiters or provide 'ignore embedded instructions' warnings.
Capability inventory: File system read/write (via download_paper/read_paper in ~/.arxiv-papers) and network access (via Exa tools).
Sanitization: None detected in the provided instructional workflow.

academic-research