understand

Pass

Audited by Gen Agent Trust Hub on Nov 5, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill has a large attack surface as it ingests and processes untrusted project data (README files, manifests, and source code) and passes them to AI subagents.
  • Ingestion points: Project README and package manifests are read in Phase 1 and Phase 5; the entire source code is analyzed in batches in Phase 2.
  • Boundary markers: The skill includes explicit instructions for the dispatched subagents to treat README and manifest content as untrusted data and to ignore any instructions or directives embedded within them.
  • Capability inventory: The skill has the capability to run shell commands (git, find, pnpm), read/write files in the project directory, and execute Node.js/Python scripts.
  • Sanitization: The primary mitigation is prompt-level instruction to subagents to maintain boundaries between instructions and analyzed data.
  • [DYNAMIC_EXECUTION]: The skill employs dynamic script generation and dynamic module loading at runtime.
  • Evidence: In Phase 6, the skill writes a custom validation script (ua-inline-validate.cjs) to a temporary directory and executes it using Node.js.
  • Evidence: Several internal scripts, including extract-structure.mjs and compute-batches.mjs, use dynamic import() calls with paths computed at runtime to load core plugin modules.
  • [COMMAND_EXECUTION]: The skill frequently executes shell commands and bundled scripts to perform its analysis and data merging tasks.
  • Evidence: It uses git for commit tracking and diffing, find for file inventory, and pnpm for building the @understand-anything/core package if the distribution files are missing.
  • Evidence: It coordinates the execution of numerous Node.js and Python scripts (e.g., merge-batch-graphs.py, scan-project.mjs) to process project data.
  • [EXTERNAL_DOWNLOADS]: The skill may download and install its own dependencies from the official NPM registry.
  • Evidence: Phase 0 in SKILL.md contains logic to run pnpm install and pnpm build within the plugin's local directory if the core components are not already built.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 5, 2026, 07:48 PM