The Agent Skills Directory

[SAFE]: A comprehensive analysis of the skill's instructions and Python scripts revealed no security vulnerabilities or malicious patterns.
[DATA_EXPOSURE]: The tools are designed to operate on local files provided by the user, such as CSV feature lists and PNG screenshots. There are no hardcoded credentials or network exfiltration capabilities.
[REMOTE_CODE_EXECUTION]: The scripts do not employ dynamic code execution methods like eval() or exec(). Execution is limited to standard Python logic for file I/O and data processing.
[OBFUSCATION]: The codebase is transparent and well-documented. There is no evidence of hidden content, multi-layer encoding, or homoglyph substitution.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data via CSV and JSON files. Ingestion points: replica/features.csv (via csv.DictReader) and replica/diffs/*.json (via json.load). Boundary markers: Data is parsed into specific schema fields. Capability inventory: Local file read/write only; no network, subprocess, or dynamic execution. Sanitization: The scripts use standard library parsing for structured data. The absence of dangerous capabilities prevents these ingestion points from being leveraged for malicious actions.

replica-diff