pr-walkthrough

Warn

Audited by Gen Agent Trust Hub on Nov 19, 2026

Risk Level: MEDIUMINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONEXTERNAL_DOWNLOADSCOMMAND_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external GitHub Pull Requests to generate visualization nodes and tour steps.
  • Ingestion points: SKILL.md fetches PR titles, bodies, and review discussions using gh pr view and the GitHub API (gh api).
  • Boundary markers: Untrusted PR content is interpolated into the graph data without the use of delimiters or instructions to ignore embedded commands.
  • Capability inventory: The skill possesses capabilities to write files to the local file system, execute Python scripts, and push commits to remote Git branches.
  • Sanitization: While the output script uses basic HTML escaping for the final visualization, the agent's internal processing of the data lacks sanitization against malicious instructions.
  • [DYNAMIC_EXECUTION]: The skill generates script-heavy HTML files from templates and then executes them during a validation phase.
  • The script validate_d3_canvas.py uses the Playwright library to launch a headless Chromium browser and navigate to the locally generated file:// URL.
  • This validation step executes JavaScript within the context of the generated file, which contains data sourced from external, untrusted PR content.
  • [EXTERNAL_DOWNLOADS]: The generated walkthrough fetches the D3 visualization library from a well-known CDN at runtime.
  • Target: https://cdn.jsdelivr.net/npm/d3@7.9.0/dist/d3.min.js (jsDelivr).
  • [COMMAND_EXECUTION]: The skill performs multiple shell operations and script executions.
  • Executes local Git commands (git diff, git log, git worktree) and GitHub CLI commands (gh pr, gh api) to gather context.
  • Runs local Python scripts (d3_canvas_runtime.py and validate_d3_canvas.py) to process data and perform browser-based validation.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 19, 2026, 01:13 AM