gs-qa-scenarios

Warn

Audited by Gen Agent Trust Hub on Nov 20, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [DYNAMIC_EXECUTION]: The skill dynamically discovers and executes local skills matching the naming pattern verify-* found within project directories such as .cursor/skills/, .claude/skills/, and .agents/skills/ during the scenario execution phase.
  • [DYNAMIC_EXECUTION]: For application setup, the skill identifies and executes arbitrary commands extracted from the project's own repository files, specifically README.md, justfile, Makefile, and package.json scripts.
  • [COMMAND_EXECUTION]: The skill makes extensive use of system shell commands including git (for revision history and diffing) and the GitHub CLI (gh) for reading pull request metadata and posting comments with attachments.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input from pull request bodies, titles, and commit messages to generate test scenarios. While it includes instructions to treat embedded instructions as descriptive data, the ingest-to-execution pipeline (where generated scenarios can later be run via shell tools) presents an indirect injection surface.
  • Ingestion points: SKILL.md (Step 1
  • reading PR title, body, and commits via gh and git).
  • Capability inventory: Shell command execution (git, gh, npx), file system writes (.verify/), and network operations (gh pr comment).
  • Boundary markers: Relies on natural language separators and document structure rather than strict escaping or delimiters.
  • Sanitization: No explicit technical sanitization is defined for external text before processing, though it includes a manual step to remove secrets and tokens before outputting results.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 20, 2026, 07:02 PM