buyer-personas
Pass
Audited by Gen Agent Trust Hub on Jul 17, 2026
Risk Level: SAFEPROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The skill is designed to process untrusted external data, specifically sales call transcripts, survey results, and CRM notes, which creates a surface for indirect prompt injection. Malicious instructions embedded within these source materials could attempt to influence the agent's output or behavior during the extraction and synthesis phases.\n
- Ingestion points: Untrusted data enters the context via Fathom transcripts, win-loss interviews, surveys, and CRM notes as described in SKILL.md and references/five-rings.md.\n
- Boundary markers: The skill uses instruction-level markers such as 'Using only the provided sources' in the extraction prompts found in references/five-rings.md, which helps mitigate but does not fully prevent the risk of following instructions embedded in the data.\n
- Capability inventory: The skill leverages Fathom MCP tools to retrieve transcript data, as noted in the 'Andrew/Pesty note' section of SKILL.md.\n
- Sanitization: The skill explicitly requires 'verbatim quotes' and discourages sanitization or summarization in the 'Faithfulness gate' section, increasing the likelihood that any injection strings in the source data would be processed directly by the model.
Audit Metadata