converse
Warn
Audited by Gen Agent Trust Hub on Aug 3, 2026
Risk Level: MEDIUMREMOTE_CODE_EXECUTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSPROMPT_INJECTION
Full Analysis
- [REMOTE_CODE_EXECUTION]: The skill recommends executing 'uvx voice-mode-install --yes', which downloads and runs a package from a public registry (PyPI) without version pinning or source verification, allowing for arbitrary code execution.
- [COMMAND_EXECUTION]: The implementation relies on multiple shell commands for tool installation and service management, including 'voicemode whisper service install' and 'voicemode kokoro install'.
- [EXTERNAL_DOWNLOADS]: The skill initiates downloads of external binaries and models for FFmpeg, Whisper, and Kokoro during its installation phase.
- [PROMPT_INJECTION]: The instructions direct the agent to read persona data from local markdown files ('~/.voicemode/voices//README.md') to define its behavior. This represents an indirect prompt injection surface as these files could contain instructions that override safety guidelines.
- Ingestion points: Files located at '
/.voicemode/voices//README.md' and '/.voicemode/voices/PERSONAS.md'. - Boundary markers: None identified; the agent is instructed to adopt the persona described in the files directly.
- Capability inventory: Access to shell execution ('uvx', 'voicemode') and local file system reads.
- Sanitization: None identified; content from persona files is treated as authoritative for character roleplay.
Audit Metadata