customer-segmentation
Pass
Audited by Gen Agent Trust Hub on Sep 25, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONPROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from external websites and user-provided data sources, creating a potential surface for indirect prompt injection.
- Ingestion points: The skill performs research by fetching content from public URLs (websites) and processes customer data provided via CSV uploads or text pastes.
- Boundary markers: The skill provides exceptionally strong boundary logic in
references/agent-security.md. It explicitly mandates that every retrieved byte must be treated as untrusted data, and instructions found inside such content must be reported as findings rather than followed. - Capability inventory: The skill has the capability to write to local state files in the
.agents/directory and can utilize Model Context Protocol (MCP) tools to retrieve live customer data. - Sanitization: Instructions require HTML-escaping for all interpolated values in markup and provide specific rules for redacting credentials and avoiding the use of protected identity categories (e.g., race, religion) in segmentation logic.
- [PROMPT_INJECTION]: A static detector flagged potential concealment of actions within the instructions.
- Evidence: The
SKILL.mdfile instructs the agent to "skip this step without comment" if specific visual rendering tools are unavailable. - Context: This is an instructional flow for conditional UI rendering (Artifacts) and does not indicate a malicious attempt to hide behavior or override safety protocols.
Audit Metadata