nature-academic-search
Pass
Audited by Gen Agent Trust Hub on Nov 3, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill fetches data from external scholarly databases and presents it to the agent, creating a surface for potential injection via malicious metadata.
- Ingestion points: Paper metadata is retrieved from CrossRef, PubMed, and arXiv APIs in the
mcp-server/sources/directory. - Boundary markers: Results are structured using JSON serialization in
mcp-server/academic_search_server.pyto provide clear delimiters for the agent. - Capability inventory: The skill allows searching and managing citations but does not include high-risk capabilities like arbitrary file writing or system command execution based on search results.
- Sanitization: The
ris_escapefunction inscripts/converters.pystrips HTML tags and normalizes whitespace for exported citation files. - [COMMAND_EXECUTION]: The
install.shscript automates environment setup by installing standard Python libraries (mcp,requests,toml,lxml) and modifying local agent configuration files (.mcp.jsonandsettings.json) to register the search server. - [DYNAMIC_EXECUTION]: The test suite in
mcp-server/tests/test_sources.pyemploys__import__for dynamic module loading. This is restricted to the test environment for mocking purposes and does not impact the skill's runtime search functionality.
Audit Metadata