smart-bidding
Pass
Audited by Gen Agent Trust Hub on Sep 17, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest untrusted data from campaign exports and external websites, creating a potential attack surface for indirect prompt injection.
- Ingestion points: Campaign names, strategy labels, pasted exports, and external URLs fetched during the product research phase (SKILL.md).
- Boundary markers: The skill contains explicit instructions in
references/agent-security.mdto treat every retrieved or pasted byte as untrusted content to be described, never as direction to be followed. - Capability inventory: The skill can fetch external URLs, read/write to the
.agents/product-context.mdfile, and generate advertising strategy recommendations. - Sanitization: Security guidelines require HTML-escaping for all interpolated values in emitted markup and prohibit placing untrusted content inside script blocks or event handlers.
- [COMMAND_EXECUTION]: The skill performs network operations to fetch external websites for research purposes.
- Evidence: Instructions in
SKILL.mddirect the agent to fetch named sites or pages to build product context. These operations are restricted by safety rules inreferences/agent-security.mdwhich forbid following URLs that originate from inside previously fetched content. - [PROMPT_INJECTION]: The skill defines "House Rules" that modify the agent's interaction model to prioritize efficiency.
- Evidence: The agent is instructed to limit user questions to a hard cap of three and to use stated assumptions for missing data rather than stopping the workflow (SKILL.md, references/house-rules.md). While these instructions override default conversational patterns, they are established as professional UX guidelines for this skill pack and do not attempt to bypass core safety filters.
Audit Metadata