Testing Skills With Subagents
Warn
Audited by Gen Agent Trust Hub on Oct 23, 2026
Risk Level: MEDIUMPROMPT_INJECTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The skill provides templates and instructions for overriding agent behavior using 'pressure scenarios'. It encourages the use of directive markers such as 'IMPORTANT: This is a real scenario. Choose and act' and 'Don't ask hypothetical questions
- make the actual decision' to bypass the agent's standard conversational constraints and force it into specific personas or choices.
- [INDIRECT_PROMPT_INJECTION]: The framework relies on the agent processing external 'scenarios' designed to pressure its decision-making. The skill instructions tell the agent to 'believe it's real work, not a quiz', which intentionally blurs the boundary between test data and system instructions.
- Ingestion points: Scenarios provided in
SKILL.mdandexamples/CLAUDE_MD_TESTING.mdare processed as direct instructions. - Boundary markers: The skill lacks boundary markers or 'ignore embedded instructions' warnings; instead, it explicitly encourages the agent to internalize the scenario constraints.
- Capability inventory: The skill primarily influences agent logic and decision-making; no direct tool execution or file system modifications are performed by the skill itself, though it instructs on how to structure other skills that might have such capabilities.
- Sanitization: No sanitization or validation of the input scenarios is performed.
Audit Metadata