referral-program
Pass
Audited by Gen Agent Trust Hub on Sep 25, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to fetch content from external websites, which could potentially contain malicious instructions intended to influence the agent's behavior. \n
- Ingestion points: The skill fetches the user's own website and competitor referral pages (Step 5c in SKILL.md).\n
- Boundary markers: The
references/house-rules.mdfile contains a global instruction (Rule 7) to treat fetched content as data and never as an instruction, but no technical delimiters are specified in the individual fetching steps inSKILL.md.\n - Capability inventory: The skill uses network fetching tools and performs local file writes to
.agents/product-context.md.\n - Sanitization: The skill relies on natural language instructions to distinguish data from commands and does not implement technical sanitization or schema validation for web content.\n- [COMMAND_EXECUTION]: The skill uses platform tools to fetch external web content and read or write to the local file system (specifically
.agents/product-context.md). While these actions support its primary function of researching business context, they involve interaction with the external environment and local files.
Audit Metadata