testing-skills-with-subagents

Pass

Audited by Gen Agent Trust Hub on Mar 29, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The skill documentation includes numerous 'pressure scenarios' that utilize imperative language and role-play markers (e.g., 'IMPORTANT: This is a real scenario', 'You must choose and act') to simulate attempts to bypass safety or process rules. These patterns are explicitly intended for testing an agent's compliance with other instructions.
  • [COMMAND_EXECUTION]: The skill provides guidelines for using standard shell utilities like ls and grep to browse and search the local skills library at ~/.claude/skills/. This is used to facilitate skill discovery and verification.
  • [PROMPT_INJECTION]: The skill involves a workflow for processing external scenarios, creating a surface for indirect prompt injection.
  • Ingestion points: Test scenarios defined in SKILL.md and examples/CLAUDE_MD_TESTING.md.
  • Boundary markers: Scenarios are primarily delimited using markdown code blocks.
  • Capability inventory: Local file system listing (ls), searching (grep), and rule enforcement logic.
  • Sanitization: No explicit sanitization or validation of the scenario text is performed prior to agent processing.
Audit Metadata
Risk Level
SAFE
Analyzed
Mar 29, 2026, 05:13 PM