own-voice-builder
Pass
Audited by Gen Agent Trust Hub on Aug 30, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
- Extensive Data Access: To build a writing voice profile, the skill analyzes a sample of 100–150 sent emails. This involves the systematic processing of personal or professional communications to extract linguistic patterns, which is a standard requirement for the skill's stated goal.
- Dynamic Capability Generation: The skill uses the
m_create_skilltool to programmatically generate new agent skills. This allows the AI to define new operational rules and behaviors based on its analysis, effectively extending the agent's feature set dynamically. - Indirect Prompt Injection Surface: The skill processes external text from emails, which serves as an ingestion point for untrusted data that could theoretically contain hidden instructions.
- Ingestion points: The skill retrieves 100–150 emails from the Microsoft 365 Sent folder via the
m365_list_emailstool (SKILL.md). - Boundary markers: The instructions explicitly direct the agent to "extract patterns, not content" and to use "synthetic examples" in the final output to maintain a boundary between the source data and the generated profile.
- Capability inventory: The skill utilizes
m365_list_emailsfor data retrieval andm_create_skillfor generating new agent functionality (SKILL.md). - Sanitization: The skill includes instructions to skip auto-generated mail and mandates that verbatim sentences or confidential details must never be copied into the generated profile.
Audit Metadata