The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/normalize_depth_paths.py performs unsafe deserialization of local data.
Evidence: The normalize_ann_file function calls pickle.load(handle) on user-provided .pkl files.
Risk: The pickle module is known to be insecure as it can execute arbitrary code during the loading process. If an attacker can provide a malicious .pkl file, they could achieve code execution on the user's system.
Remediation: Replace pickle with a safer format like JSON for annotation files, or use a library that provides secure deserialization.
[INDIRECT_PROMPT_INJECTION]: The skill has an attack surface for indirect prompt injection through processed data.
Ingestion points: The scripts/normalize_depth_paths.py script ingests data from .pkl files provided via command-line arguments.
Boundary markers: The instructions and processing script do not use boundary markers or "ignore embedded instructions" warnings for the processed annotation content.
Capability inventory: The script has the ability to write files to the filesystem using pickle.dump() in scripts/normalize_depth_paths.py.
Sanitization: There is no validation or sanitization of the contents of the annotation files before they are processed and written back.
Remediation: Implement strict schema validation for all ingested data and add delimiters when interpolating data into prompts.

tao-train-sparse4d