academic-research
Pass
Audited by Gen Agent Trust Hub on Nov 1, 2026
Risk Level: SAFEEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill documentation recommends installing a third-party MCP server (
arxiv-mcp-server) from an unofficial GitHub repository (blazickjp/arxiv-mcp-server) usinguvornpx. - [DATA_EXPOSURE]: The
SKILL.mdfile contains a hardcoded reference to a specific local directory path:~/Desktop/Aldea/01-Repos/perplexity-clone/model/moba_block_sparse.py. This reveals information about the user's local filesystem structure. - [COMMAND_EXECUTION]: The skill executes a custom bash script
academic_stealth_fetch.shwhich invokes theobscuraCLI tool to perform web scraping operations. - [DYNAMIC_EXECUTION]: The
academic_stealth_fetch.shscript dynamically constructs JavaScript strings and passes them to theobscuratool's--evalflag for execution within a headless browser environment. - [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external academic sources which could contain malicious instructions.
- Ingestion points: The skill ingests search results from Exa, paper content via
read_paper(ArXiv MCP), and scraped metadata fromobscura. - Boundary markers: No specific delimiting or safety instructions are provided to the agent for handling these external data streams.
- Capability inventory: The agent has the ability to search web content, download and read files from the ArXiv repository, and execute shell scripts.
- Sanitization: Although the scraping script uses
JSON.stringifyfor data formatting, no content filtering or instruction-stripping is performed before the data is presented to the agent's context.
Audit Metadata