explain-changes

Warn

Audited by Gen Agent Trust Hub on Nov 12, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
  • [COMMAND_EXECUTION]: The script assets/review.py uses subprocess.run with shell=True in several places. Specifically, the begin and render commands execute user-provided shell strings via the --capture argument to record system state (e.g., crontab -l). The run command also executes arbitrary shell commands provided as arguments. This is the primary functionality of the tool but poses a risk if an attacker can influence the command strings.
  • [DYNAMIC_EXECUTION]: The skill facilitates the creation and rendering of HTML/JS files (review.html) that incorporate script results and file diffs. The script assets/review.py inlines assets/review.js and assets/review.css into the generated report. Although these scripts are local, they handle data from git and system commands which could be manipulated.
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to process external data (git diffs, file contents, command outputs) and interpolate them into a report for the agent to summarize. There is a risk that malicious instructions embedded in a processed file or a command's output could influence the agent's summary or the user's perception of the changes.
  • Ingestion points: assets/review.py reads file contents via Path.read_bytes(), git diffs via subprocess.run(['git', 'diff', ...]), and command outputs via run_capture.
  • Boundary markers: The skill uses markdown headers and HTML structure to separate content, but lacks explicit 'ignore embedded instructions' warnings in the generated summaries.
  • Capability inventory: Includes file reading, file writing (rendering the report), and arbitrary shell command execution via the run and begin --capture commands.
  • Sanitization: Content is escaped using html.escape before being placed in the HTML report, which mitigates XSS, but does not prevent the agent from obeying instructions found within the data it is asked to summarize.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 12, 2026, 03:36 PM