voicebox
Pass
Audited by Gen Agent Trust Hub on Sep 3, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill defines an attack surface by reading user dictation history which could contain malicious instructions. 1. Ingestion points: Dictation transcripts and captures retrieved via the voicebox_list_captures tool (SKILL.md). 2. Boundary markers: Absent in the instructions. 3. Capability inventory: The agent can synthesize audio using voicebox_speak but lacks high-risk write or network capabilities. 4. Sanitization: Not explicitly defined in the skill logic.
- [EXTERNAL_DOWNLOADS]: The documentation indicates that the Voicebox application downloads Whisper models locally upon first use. This is a legitimate functional requirement of the external software and is not executed as a script from the skill itself.
Audit Metadata