voice
Pass
Audited by Gen Agent Trust Hub on Oct 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted data from external users, such as survey responses, reviews, and support tickets, creating a potential surface for indirect prompt injection.
- Ingestion points: Untrusted content enters the agent context through feedback collection channels (NPS, CSAT, CES, reviews) and is processed by subcommands like
review,sentiment, andthematic(SKILL.md, reference/multi-channel-synthesis.md). - Boundary markers: The instructions specify the use of structured output schemas and 'contamination gates' (synthetic feedback detection) to filter inputs before they are processed by LLM pipelines.
- Capability inventory: The skill performs sentiment analysis, classification, and thematic mapping. Findings are routed to other specialized agents (Spark, Growth, Compete) for further action. There is no evidence of the skill executing arbitrary shell commands or writing to sensitive file paths.
- Sanitization: The skill employs 'Model Variability Problem' (MVP) mitigations, including multi-run ensemble consensus and entropy-based uncertainty quantification to validate outputs derived from untrusted text.
- [EXTERNAL_DOWNLOADS]: The documentation references several well-known technology platforms, research organizations, and regulatory sources for benchmarking and industry context.
- Evidence: References include Gartner, Retently, CustomerGauge, White & Case, and specialized feedback tools like Hotjar, Qualtrics, and Medallia.
- Context: These references are informational and point to reputable sources to support the agent's analytical logic (e.g., 2026 NPS benchmarks, EU AI Act compliance guidelines). They do not involve the execution of remote code.
Audit Metadata