p-literature-download

Pass

Audited by Gen Agent Trust Hub on Sep 23, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill executes external processes to launch and manage the automation environment.
  • scripts/edge_download.py uses subprocess.run to execute a PowerShell command block that decrypts WebVPN credentials using the Import-Clixml cmdlet and Windows DPAPI.
  • scripts/start_edge.ps1 uses Start-Process to launch Microsoft Edge with specific debugging flags (--remote-debugging-port) and profile settings.
  • [EXTERNAL_DOWNLOADS]: The skill facilitates the retrieval of PDF documents and metadata from remote sources related to its primary purpose.
  • Fetches academic papers from various publisher domains (e.g., ScienceDirect, Wiley, Nature, ACS) using Playwright and standard HTTP requests.
  • Retrieves publication metadata from api.crossref.org and eutils.ncbi.nlm.nih.gov (NLM Catalog) to verify and name files.
  • [DYNAMIC_EXECUTION]: The skill uses browser automation to interact with dynamic web content.
  • Injects JavaScript snippets via page.evaluate() to perform DOM scraping, handle cookie consent banners, and extract PDF data as Base64-encoded chunks from the browser's memory (JS_FETCH_PDF, JS_FIND_PDF_URLS).
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from academic publisher websites and metadata APIs.
  • Ingestion points: Data is ingested through browser page content (HTML) and metadata API responses (scripts/literature_download.py).
  • Boundary markers: None identified in the parsing logic for external metadata.
  • Capability inventory: The skill can write files (PDFs, reports, credentials), make network requests, and execute shell commands via PowerShell.
  • Sanitization: The skill includes verification logic in scripts/verify_pdf.py which checks PDF integrity (header, trailer, page count) and verifies that the content matches the requested DOI or title, providing a layer of validation for ingested data.
  • [OBFUSCATION]: The skill uses Base64 encoding in scripts/edge_download.py (fetch_in_page) to transfer binary PDF data from the browser context to the Python execution environment. This is a common data-passing pattern in Playwright/CDP automation and is not used to hide malicious code.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 23, 2026, 07:50 AM