21-ads-audit-global
Pass
Audited by Gen Agent Trust Hub on Nov 7, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes external data from ad account exports, dashboards, and live platform connectors (MCP), creating a surface for indirect prompt injection.
- Ingestion points: Data enters the skill context via text tables, dashboard links, and Model Context Protocol (MCP) servers.
- Boundary markers: The skill uses a structured evaluation framework, although it does not implement explicit text-based delimiters for all ingested content.
- Capability inventory: The skill has read/write access to Meta, Google, and TikTok advertising platforms through connected MCP servers.
- Sanitization: The skill strongly mitigates potential injection risks by enforcing a 'draft-first' policy (Rule 7), ensuring that any recommended changes are presented as approval blocks rather than being executed automatically.
- [COMMAND_EXECUTION]: The skill leverages MCP tools to perform platform operations. It explicitly notes that tools such as the Meta Ads MCP have irreversible write capabilities and requires a human-in-the-loop approval process for any action that modifies a live account.
- [EXTERNAL_DOWNLOADS]: The skill refers to official platform MCPs (Meta, Google) and community-maintained servers (TikTok Ads MCP, Adspirer). These are documented as integration points for the user and do not involve hidden or malicious remote code execution.
Audit Metadata