The Agent Skills Directory

[PROMPT_INJECTION]: The skill instructions and seed patterns contain no attempts to override agent behavior, bypass safety filters, or extract system prompts. The use of 'important' in documentation is strictly for guiding data processing logic.
[DATA_EXFILTRATION]: No network activity or exfiltration patterns were detected. All data processing is performed locally using pandas and scikit-learn, and file outputs are restricted to the local workspace.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data (CSV, JSONL, Parquet, Excel) which represents a potential attack surface. However, the risk is mitigated by:
Ingestion points: Data is loaded through standard pandas readers in detect_issues.py and handle_missing.py.
Sanitization: normalize_strings.py explicitly removes Unicode control characters and applies NFC normalization, which helps neutralize common injection payloads.
Capability inventory: The skill's operations are restricted to data mutation and file-system writes to new files.
Boundary markers: The instructions explicitly warn the agent to verify whether 'null-like' strings (e.g., 'none', 'null') are legitimate content or actual missing values, reducing the likelihood of accidental data corruption.
[REMOTE_CODE_EXECUTION]: The skill does not download external scripts or execute code from untrusted remote sources. It relies on standard environment dependencies (pandas, numpy, scikit-learn).
[COMMAND_EXECUTION]: File operations and cleaning tasks are executed through structured Python scripts. There is no evidence of arbitrary shell command execution or unsafe argument interpolation into subprocesses.
[OBFUSCATION]: The skill includes character mapping for 'mojibake' (e.g., Ã© to é) and Unicode normalization. These are legitimate data cleaning operations and do not represent malicious obfuscation.

magic-data-cleaning