google-ads-change-plan
Pass
Audited by Gen Agent Trust Hub on Sep 17, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external data, which creates a surface for indirect prompt injection attacks.
- Ingestion points: Untrusted data enters the agent context through CSV/export files, Google Ads findings Produced by other skills, and content fetched from external websites or pages as specified in
SKILL.mdandreferences/agent-security.md. - Boundary markers: The skill body and the
references/agent-security.mdfile contain robust instructions to treat retrieved content strictly as data and never as instructions. It specifically commands the agent to quote rather than summarize potential injection attempts to prevent instruction propagation. - Capability inventory: The skill has the capability to write research findings to
.agents/product-context.mdand may use visual artifact tools if available on the platform. - Sanitization:
references/agent-security.mdmandates strict sanitization protocols, including HTML escaping of all interpolated values, validation of URLs to HTTPS or relative paths, and redaction of credentials or tokens. - [SAFE]: A heuristic detector flagged potential concealment in
SKILL.md, but analysis confirms these instructions are benign operational guidelines. The instruction to ask fewer questions and avoid producing 'drafts with warnings' is intended to prevent the generation of low-quality or incomplete outputs, serving as a quality control measure rather than a malicious attempt to hide actions from the user. Similarly, the instruction to handle missing context files by researching and persisting them is a standard context-management feature.
Audit Metadata