daily-papers

Fail

Audited by Gen Agent Trust Hub on Oct 27, 2026

Risk Level: HIGHCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [COMMAND_EXECUTION]: The script enrich_papers.py uses asyncio.create_subprocess_shell to execute a command pipeline that includes an unvalidated arxiv_id parameter. This parameter is sourced from input JSON data which can be manipulated to include shell metacharacters (e.g., ;, &, |), leading to arbitrary command execution on the host system.
  • [REMOTE_CODE_EXECUTION]: The skill implements a execution pipeline that pipes output from curl and pdftotext into a Python process. While intended to parse PDF content, the lack of input sanitization in the shell command construction creates a high-severity execution risk.
  • [DYNAMIC_EXECUTION]: In enrich_papers.py, the pathlib module is loaded using the __import__ function rather than a standard import statement. This dynamic execution pattern is an unusual choice that can be used to evade static analysis of imports.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to fetch and process titles and abstracts from arXiv and HuggingFace to generate recommendations and notes. This untrusted content is ingested into the agent's context without sanitization, creating a surface for indirect prompt injection attacks.
  • Ingestion points: Data is fetched from HuggingFace API and arXiv API in fetch_and_score.py, and from arXiv HTML/PDF pages in enrich_papers.py and download_note_images.py.
  • Boundary markers: The scripts do not implement delimiters or safety instructions to distinguish between paper content and agent instructions.
  • Capability inventory: The skill has the ability to execute shell commands (create_subprocess_shell), perform network requests (curl), and write files to the local system.
  • Sanitization: No validation or sanitization is performed on the retrieved academic content before it is processed or presented to the agent.
  • [EXTERNAL_DOWNLOADS]: The skill fetches paper metadata and PDF content from HuggingFace and arXiv. While these are well-known academic repositories, the automated nature of the downloads combined with the command execution risks increases the overall attack surface.
Recommendations
  • HIGH: Downloads and executes remote code from: https://arxiv.org/pdf/{arxiv_id}, https://export.arxiv.org/api/query?... - DO NOT USE without thorough review
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Oct 27, 2026, 03:40 PM