oma-voice
Pass
Audited by Gen Agent Trust Hub on Oct 6, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill ingests and processes untrusted audio data for transcription, creating a vector for indirect prompt injection where malicious instructions are embedded in audio recordings.
- Ingestion points: The skill reads local audio files (mp3, wav, m4a, etc.) as described in SKILL.md and execution-protocol.md to perform speech-to-text operations.
- Boundary markers: No delimiters or safety instructions are provided to the agent to distinguish between transcribed content and system instructions in the resulting transcript.md output.
- Capability inventory: The skill has file-write permissions for the results directory and makes network requests to the loopback address 127.0.0.1:17493.
- Sanitization: There is no automated sanitization or filtering of transcribed text to remove potential injection patterns before the content is presented to the agent.
Audit Metadata