customer-research
Pass
Audited by Gen Agent Trust Hub on Oct 12, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from various sources including customer transcripts, support tickets, and public community posts (Reddit, G2, LinkedIn, etc.). This data is ingested into the agent's context for analysis and synthesis. The absence of specific boundary markers or instructions to disregard embedded commands within these data sources presents a surface for indirect prompt injection.
- Ingestion points: Processes research assets like transcripts and surveys in
SKILL.mdand gathers intelligence from digital platforms like Reddit and review sites as described inreferences/source-guides.md. - Boundary markers: There are no instructions in the skill files to use delimiters or ignore potentially malicious instructions embedded within the ingested research material.
- Capability inventory: The skill focus is on data extraction, clustering, and persona generation; it does not explicitly call for high-risk operations like shell execution or file system writes based on the external input, though the underlying agent platform may possess such capabilities.
- Sanitization: No sanitization, validation, or filtering steps are defined for the external research data before it is processed by the agent.
Audit Metadata