The Agent Skills Directory

[DYNAMIC_EXECUTION]: Multiple scripts within the scripts/ directory (such as 02_path_empty.py, 04_drone_attitude.py, and 06_night.py) use the exec(open(...).read()) pattern to dynamically load and run a helper library (tools/lib.py). The file path is determined at runtime based on environment variables, which is a significant dynamic execution vector.
[REMOTE_CODE_EXECUTION]: The tools/bridge.py script starts a TCP server listening on 127.0.0.1:9876. This server accepts arbitrary Python code and executes it using exec(compile(code, ...)). Although restricted to the local interface, this creates a local unauthenticated remote code execution (RCE) surface.
[COMMAND_EXECUTION]: The skill uses subprocess.run in scripts/12_build_sfx.py and a shell script in tools/clay_assemble.sh to execute ffmpeg commands for audio and video processing. This involves spawning external system processes.
[INDIRECT_PROMPT_INJECTION]: The skill's workflow involves Claude generating storyboards and motion options which are then used as parameters for the Blender scripts. This represents an attack surface where untrusted or maliciously influenced data could potentially affect the execution path, although the current implementation processes these as structured data (JSON).
Ingestion points: scripts/10_render_edit.py reads from command line arguments; tools/bridge.py receives code over a network socket; scripts load data from generated JSON files (e.g., research/sim_out.json).
Boundary markers: None identified in the prompt interpolation or data processing scripts.
Capability inventory: The skill has extensive capabilities including arbitrary Python execution (exec), file system access, and external command execution via subprocesses.
Sanitization: There is no significant sanitization of the code or data processed by the dynamic execution or bridge components.

blender-film