testing-skills-with-subagents
Pass
Audited by Gen Agent Trust Hub on Sep 17, 2026
Risk Level: SAFEPROMPT_INJECTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The skill body and examples include numerous prompt injection strings used as test cases, such as "IMPORTANT: This is a real scenario. Choose and act." and "Ignore previous instructions". Because these are presented as plain text without escaping or specific markers, they could be interpreted by an AI agent as direct commands rather than illustrative examples.\n- [INDIRECT_PROMPT_INJECTION]: The skill facilitates the evaluation of agent behavior by ingesting untrusted "pressure scenarios" that simulate adversarial prompts.\n
- Ingestion points: The pressure scenarios located in
SKILL.mdandexamples/CLAUDE_MD_TESTING.md.\n - Boundary markers: The skill fails to provide or recommend boundary markers (such as triple backticks or explicit ignore directives) for these scenarios, increasing the risk that an agent may treat the scenario content as part of its operational instructions.\n
- Capability inventory: The skill provides instructions that influence the agent's logic for skill discovery and adherence, which governs how it approaches all other tasks.\n
- Sanitization: There is no mention of sanitizing, escaping, or validating the content of these test scenarios before they are presented to the agent or subagent.
Audit Metadata