qwen-subtitle
Warn
Audited by Gen Agent Trust Hub on Aug 7, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONCREDENTIALS_UNSAFEPROMPT_INJECTIONDATA_EXFILTRATIONEXTERNAL_DOWNLOADS
Full Analysis
- [COMMAND_EXECUTION]: The scripts
scripts/subfix.pyandscripts/dub_multi.pyutilizesubprocess.runto execute external binaries includingffmpeg,ffprobe, and thebl(Bailian) CLI for video processing and model interaction. - [CREDENTIALS_UNSAFE]: The
scripts/dub_multi.pyscript contains logic to read theapi_keydirectly from~/.bailian/config.json. While intended for service authentication, accessing credentials from fixed filesystem paths is a sensitive operation. - [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection via the media content it processes.
- Ingestion points:
scripts/subfix.py(ingests video frames viabl vision describeand ASR text viabl text chat). - Boundary markers: Absent; the instructions for the LLM do not use strict delimiters to separate instructions from the data to be corrected.
- Capability inventory:
subprocess.runcalls for media manipulation and CLI interactions insubfix.pyanddub_multi.py. - Sanitization: Absent; the output from the models is directly used to generate subtitle files without secondary validation.
- [DATA_EXFILTRATION]: Audio tracks and video frames are extracted locally and uploaded to Alibaba Cloud's DashScope/Bailian services for AI processing. This is a functional requirement but involves transmitting local data to a third-party cloud provider.
- [EXTERNAL_DOWNLOADS]: The skill depends on the installation of the
blCLI tool from Alibaba Cloud and theflaskPython package for its preview functionality. - [NETWORK_EXPOSURE]: The
scripts/preview_editor.pyscript starts a Flask server binding to0.0.0.0. This makes the local video preview and subtitle editor accessible to any device on the same local network without authentication.
Audit Metadata