coze-voice-gen

Pass

Audited by Gen Agent Trust Hub on Aug 11, 2026

Risk Level: SAFEDATA_EXFILTRATIONPROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [DATA_EXFILTRATION]: The scripts/asr.ts script uses fs.readFileSync to read local files specified via the --file command-line argument. The resulting data is converted to base64 and transmitted to an external API via the ASRClient. While intended for audio processing, this mechanism could be used to read and transmit sensitive local files if the agent is directed to process non-audio paths.
  • [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection through its Speech-to-Text (ASR) feature. Transcribed text from audio files is printed directly to stdout and subsequently ingested by the agent. If the audio source contains malicious instructions, the agent may execute them as there are no boundary markers or sanitization processes.
  • Ingestion points: Transcribed text output in scripts/asr.ts.
  • Boundary markers: None.
  • Capability inventory: Local file system access (fs.readFileSync), network communication via the external SDK clients.
  • Sanitization: None.
  • [EXTERNAL_DOWNLOADS]: The skill depends on the coze-coding-dev-sdk package to perform audio synthesis and recognition, which involves sending data to and receiving data from external service endpoints.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 11, 2026, 05:47 AM
Security Audit — agent-trust-hub — coze-voice-gen