creating-skills

Fail

Audited by Gen Agent Trust Hub on Feb 16, 2026

Risk Level: HIGHPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION] (HIGH): The file references/skill-activation.md provides explicit strategies for overriding an AI agent's reasoning.\n
  • Evidence: The document recommends using high-pressure keywords like 'MANDATORY', 'CRITICAL', and 'WORTHLESS' to force the agent into a 'Commitment Mechanism'.\n
  • Evidence: It describes 'Forced eval hooks' designed to compel the AI to 'show its work' (YES/NO evaluation) and commit to activation, bypassing its default implementation logic.\n
  • Impact: This constitutes a direct attempt to override agent behavior and bypass the AI's standard operational safety and decision-making filters by leveraging linguistic pressure.\n- [COMMAND_EXECUTION] (LOW): The scripts scripts/package_skill.py and scripts/verify.py perform file system operations.\n
  • Evidence: package_skill.py performs directory traversal and zip archive creation.\n
  • Evidence: verify.py reads and parses local file content (SKILL.md) using regular expressions.\n
  • Context: These are functional behaviors for packaging and validation, but they demonstrate the skill's capability to interact with the local host environment.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Feb 16, 2026, 08:47 AM