The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/normalize_depth_paths.py uses the pickle.load() function to read annotation data from files provided as command-line arguments. In Python, the pickle module is insecure and can execute arbitrary code during deserialization. If an attacker provides a maliciously crafted pickle file (e.g., as part of a dataset), running this script would compromise the environment.
[COMMAND_EXECUTION]: The skill makes extensive use of the Bash tool to execute model training, conversion, and quantization commands (e.g., sparse4d train, annotations convert). These commands involve interpolating user-provided paths and configuration values. Without strict validation, this pattern presents a potential risk for command injection or execution of unauthorized shell operations.
[INDIRECT_PROMPT_INJECTION]: The skill has a significant attack surface because it is designed to ingest and process external, potentially untrusted datasets and configuration files.
Ingestion points: Annotation metadata files (.pkl) and H5 files processed by the normalization script.
Boundary markers: No specific delimiters or "ignore instructions" warnings are used when processing these external data files.
Capability inventory: The skill can perform arbitrary shell commands via the Bash tool and write files to the system.
Sanitization: The provided Python utility does not perform sanitization or integrity checks on the pickle data before loading it into memory.

tao-train-sparse4d