The Agent Skills Directory

[Data Exposure & Exfiltration] (SAFE): The skill states it has no network access, and manual review of the script confirms no network-related libraries (e.g., requests, urllib) or commands (e.g., curl, wget) are present. It only reads workspace metadata and writes a report.
[Remote Code Execution] (SAFE): No remote code execution patterns were detected. The script uses standard file system operations and a safe YAML loader.
[Dynamic Execution] (SAFE): The script uses yaml.safe_load() to process pipeline metadata, preventing unsafe deserialization attacks. No use of eval(), exec(), or subprocess spawning with untrusted input was found.
[Indirect Prompt Injection] (LOW): The skill ingests data from UNITS.csv and pipeline specification files. While this is an untrusted data surface, the skill only uses this data to verify file existence on the local disk. The output is a markdown report that uses backticks to delimit file paths and IDs, minimizing the risk of downstream interpretation issues.

artifact-contract-auditor