math-modeling

Warn

Audited by Gen Agent Trust Hub on Sep 12, 2026

Risk Level: MEDIUMDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [DYNAMIC_EXECUTION]: The script tools/docx/scripts/accept_changes.py dynamically generates a LibreOffice Basic macro and writes it to a temporary user profile to automate the acceptance of tracked changes in Word documents.
  • [DYNAMIC_EXECUTION]: The DeepSeek Harness plugin dsh-plugin/math-modeling-agent/plugins/math-modeling.js dynamically registers an AI agent skill by reading the contents of SKILL.md from the local file system at runtime.
  • [INDIRECT_PROMPT_INJECTION]: The skill has a large attack surface for indirect prompt injection. * Ingestion points: Processes untrusted external data from PDF (tools/pdf/scripts/), Excel (tools/xlsx/scripts/), and CSV files. * Boundary markers: While the skill provides structured templates for reports and papers, it does not explicitly use delimiters or "ignore instructions" markers when interpolating extracted data into agent context. * Capability inventory: The skill possesses extensive capabilities including arbitrary shell command execution via subprocess in multiple scripts (e.g., tools/latex/scripts/latex_paper.py), file writing to the project workspace, and network access to academic search APIs. * Sanitization: No explicit sanitization or filtering of contents from external files was observed before processing.
  • [COMMAND_EXECUTION]: The skill frequently executes external system commands using the subprocess module. This is used to invoke soffice (LibreOffice), pandoc, xelatex, latexmk, and various PDF utilities (pdfimages, pdftoppm).
  • [EXTERNAL_DOWNLOADS]: The skill performs network operations to fetch academic metadata and search results from api.openalex.org and api.anysearch.com. It also references GitHub for star history visualization.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Sep 12, 2026, 03:43 PM
Security Audit — agent-trust-hub — math-modeling