The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run in scripts/postprocess.py to call ffmpeg and ffprobe for video post-processing tasks, including metadata probing, audio muxing, and subtitle burning. While commands are constructed using argument lists, the binary paths and target file locations are configurable via config.yaml and user input.
[DATA_EXFILTRATION]: In scripts/wan_client.py, the _encode_local_file function reads arbitrary local file paths provided in the reference_images and reference_videos fields of the input JSON. These files are Base64 encoded and transmitted to the external DashScope API (dashscope.aliyuncs.com) as media payloads. This allows for the reading and transmission of local files to a remote service if the agent is directed to process sensitive paths.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because user-provided text (title, subtitle, prompt, highlights) is interpolated directly into the prompt templates in scripts/prompt_builder.py without sanitization or boundary markers. This content is subsequently sent to downstream generative models, potentially allowing malicious instructions in the input data to influence model output.
Ingestion points: User-provided JSON fields (title, subtitle, prompt, highlights) processed in scripts/generate.py.
Boundary markers: None; text is directly concatenated into multi-line prompt strings.
Capability inventory: Network requests (requests), file read/write (Path), and subprocess execution (ffmpeg).
Sanitization: None; no escaping or validation is performed on the text content before interpolation.

wan-cover-plus