testing-skills-with-subagents

Fail

Audited by Gen Agent Trust Hub on Feb 16, 2026

Risk Level: HIGHPROMPT_INJECTION
Full Analysis
  • [PROMPT_INJECTION] (HIGH): Variant C uses emphatic language ('EXTREMELY IMPORTANT') and failure threats ('you failed') to coerce the agent into a rigid behavioral pattern, bypassing standard reasoning.
  • [PROMPT_INJECTION] (HIGH): The 'Test Scenarios' section employs role-play and high-pressure context (e.g., 'Production is bleeding money') to manipulate the agent's decision-making process.
  • [PROMPT_INJECTION] (MEDIUM): Variants B and D define mandatory workflows that force the agent to query the filesystem ('ls ~/.claude/skills/') before any task, potentially overriding more efficient or safer defaults.
Recommendations
  • AI detected serious security threats
Audit Metadata
Risk Level
HIGH
Analyzed
Feb 16, 2026, 08:42 AM