audience-personas
Pass
Audited by Gen Agent Trust Hub on Aug 23, 2026
Risk Level: SAFEPROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data, specifically customer interviews, reviews, and CRM notes provided as input. This creates an attack surface where instructions hidden within the customer data could attempt to influence the agent's behavior. The skill includes instructions to keep observed language distinct from suggested copy and to flag claims for review, which helps mitigate accidental interpretation of data as instructions. Crucially, the skill lacks access to the network or sensitive file system operations, which minimizes the risk of data exfiltration or system compromise.
- Ingestion points: User-provided customer evidence (interviews, comments, reviews, CRM notes) in SKILL.md.
- Boundary markers: The instructions mandate keeping direct quotes separate from summaries and observed language distinct from suggested copy, though explicit programmatic delimiters are not defined.
- Capability inventory: No subprocess calls, file-write operations, or network requests are present in the skill instructions.
- Sanitization: No programmatic sanitization is performed, though the skill logic requires labeling confidence and flagging claims for review.
Audit Metadata