content-case-study
Pass
Audited by Gen Agent Trust Hub on Oct 31, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes external data from customer interviews, creating a standard surface for indirect prompt injection. (1) Ingestion points: Customer responses from
references/customer-interview.mdare integrated into the final output. (2) Boundary markers: Explicit rules for verbatim quotes and a required written consent check are present. (3) Capability inventory: The skill does not have dangerous command execution, system writing, or network capabilities. (4) Sanitization: The instructions mandate stripping GPS metadata from photos and redacting PII like street addresses and license plates. - [DATA_EXFILTRATION]: The skill manages customer stories and photos but enforces privacy protections, such as neighborhood-only location data and EXIF metadata removal, to mitigate unauthorized data exposure risks.
Audit Metadata