skills/garrytan/gstack/design-review/Gen Agent Trust Hub

design-review

Warn

Audited by Gen Agent Trust Hub on Nov 18, 2026

Risk Level: MEDIUMEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONDYNAMIC_EXECUTIONOBFUSCATIONDATA_EXFILTRATIONPRIVILEGE_ESCALATION
Full Analysis
  • [EXTERNAL_DOWNLOADS]: The skill contains logic to download and install the 'impeccable' design detector engine from GitHub releases if the user consents. Evidence found in SKILL.md under the 'Install offer' section: bun ... run ... gstack-design-detect.ts install --host claude which targets github.com/pbakaus/impeccable.
  • [COMMAND_EXECUTION]: The skill frequently executes shell commands to probe the environment, manage git states, and drive browser automation. Example: ~/.claude/skills/gstack/bin/gstack-skill-start, git status --porcelain, and various aside repl calls.
  • [DYNAMIC_EXECUTION]: The skill generates and executes JavaScript snippets at runtime to drive the 'Aside' browser or a headless fallback. Evidence: aside repl '...' blocks throughout SKILL.md where scripts are dynamically constructed with variable interpolation (e.g., <url>).
  • [OBFUSCATION]: The BROWSER SETUP section in SKILL.md uses a complex shell function _gs_d() with fallback logic (gtimeout, timeout, perl) and conditional pathing that obscures simple execution flow. Additionally, several scripts are spliced from external files (dom-dump.js) into string literals for aside repl.
  • [DATA_EXFILTRATION]: While the skill claims data stays local, it captures screenshots and DOM dumps of potentially authenticated pages (via the Aside browser session). It sends telemetry via gstack-skill-end including session IDs and error messages. It also includes logic for 'outside voices' which sends source code context to external models (Codex) for review.
  • [PRIVILEGE_ESCALATION]: The skill checks for sudo and chmod capabilities in various setup scripts. It explicitly instructs the use of chmod 700 and chmod 600 on temporary directories and sensitive files like DOM dumps to manage permissions.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 18, 2026, 08:54 AM