coze-asr
Warn
Audited by Gen Agent Trust Hub on Aug 11, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The script
scripts/asr.mjsutilizes thejitilibrary to dynamically import and execute code from a path (../../../src/skill-cli.ts) that resides outside of the skill's directory. This pattern of executing code via relative path traversal depends on the specific directory structure of the host environment. - [EXTERNAL_DOWNLOADS]: The skill is designed to fetch audio data from remote URLs provided via the
--urlor-ucommand-line arguments. - [DATA_EXFILTRATION]: The skill reads local files specified through the
--fileor-farguments and uploads their contents to the Coze ASR API. While this is the intended functionality, it establishes a data path from the local file system to an external service. - [PROMPT_INJECTION]: The skill presents a surface for indirect prompt injection by processing external data from audio files or URLs. The resulting transcription is returned to the agent without explicit sanitization or boundary markers.
- Ingestion points:
--urland--filearguments inSKILL.md. - Boundary markers: Absent. There are no instructions for the agent to treat the transcribed output as untrusted content.
- Capability inventory: The skill script has the ability to read local files and initiate network connections.
- Sanitization: Absent. The raw text recognized from the audio is printed directly to standard output.
Audit Metadata