skill-comply

Pass

Audited by Gen Agent Trust Hub on Oct 14, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDYNAMIC_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted markdown files (skills or rules) as input to generate behavioral specs and test scenarios. Maliciously crafted input files could potentially influence the LLM to generate harmful setup commands or adversarial prompts during the scenario generation phase.\n
  • Ingestion points: scripts/run.py reads the target skill file provided in the command line argument.\n
  • Boundary markers: The prompt templates in prompts/scenario_generator.md and prompts/spec_generator.md use --- delimiters to wrap external skill content but do not include explicit instructions for the LLM to ignore or treat instructions within that content as data only.\n
  • Capability inventory: scripts/runner.py uses subprocess.run to execute generated setup_commands and invokes the claude CLI with generated prompts.\n
  • Sanitization: The tool uses shlex.split to parse generated commands and includes logic in _safe_sandbox_dir to prevent directory traversal outside of the intended sandbox base path.\n- [COMMAND_EXECUTION]: scripts/runner.py executes shell commands (setup_commands) that are produced by the LLM during scenario generation. While these are executed within a restricted sandbox, the commands themselves are derived from analysis of untrusted input files.\n- [DYNAMIC_EXECUTION]: The tool dynamically constructs and executes behavioral tests and environment setup steps at runtime based on the interpretation of provided skill files.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 14, 2026, 02:58 PM