The Agent Skills Directory

[DATA_EXFILTRATION]: The skill uploads user-provided media files to Volcengine TOS (TOS_BUCKET) and explicitly modifies the Access Control List (ACL) to allow public read access for all users.
Evidence: In scripts/transcribe.py, the set_public_access function sends an ACL XML configuration to the cloud storage endpoint that grants READ permission to the AllUsers group.
[COMMAND_EXECUTION]: The script performs automated video editing by executing ffmpeg as a subprocess.
Evidence: scripts/transcribe.py dynamically constructs a list of arguments for ffmpeg and executes it via subprocess.run(cmd, ...) to render the final edited video.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted transcription data (ASR results) which is then used by the agent to make logic-based decisions about video editing (e.g., determining which segments to KEEP or DELETE).
Ingestion points: Audio and video content is transcribed into text via scripts/transcribe.py.
Boundary markers: None identified in the processing logic to distinguish between data and instructions.
Capability inventory: Local file system access, network upload to Volcengine, and ffmpeg command execution.
Sanitization: No sanitization or filtering is performed on the transcribed text before it is presented to the agent for analysis.

yichen-volc-asr