ab-test
Pass
Audited by Gen Agent Trust Hub on Sep 17, 2026
Risk Level: SAFEPROMPT_INJECTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The skill contains instructions that attempt to suppress warnings or caveats in the output (‐Do not return a draft with a warning on it. The user copies the draft and leaves the warning behind, so a caveat protects you and not them.‐). Additionally, it directs the agent to skip certain tool-related steps ‐without comment‐ if a tool is missing. These instructions represent an override of default agent transparency and safety reporting behaviors.
- [INDIRECT_PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection by fetching content from external websites and public sources. This data is then written to a persistent context file (‐.agents/product-context.md‐) used by this and potentially other skills.
- Ingestion points: The agent is instructed to ‐fetch the site or page‐ named by the user and ‐research the company yourself: their site... plus public sources.‐
- Boundary markers: While the skill references a separate reference file (‐references/house-rules.md‐) that contains a general rule to treat fetched content as data, the skill itself does not employ specific delimiters or unique instructions to ignore embedded commands within the fetched content.
- Capability inventory: The skill requires network reading capabilities to fetch external URLs and file-writing capabilities to update the ‐.agents/product-context.md‐ file.
- Sanitization: The skill lacks explicit instructions to sanitize, escape, or filter the external content before it is processed and stored in the persistent context file.
Audit Metadata