The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes a Makefile and a shell entrypoint (run.sh) that execute various system commands. These include uv run for executing Python scripts, python -m honker for background workers, and litestream for database replication and restoration.
[REMOTE_CODE_EXECUTION]: The application uses joblib.load() to deserialize machine learning models in the inference path. While the artifacts are generated locally by the training worker, joblib uses the pickle protocol, which is an unsafe deserialization format. If an attacker were able to modify the model artifacts on disk or manipulate the path in the SQLite models table, it could lead to arbitrary code execution.

data-flywheel