The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted training data to perform label auditing, creating a potential surface for indirect prompt injection attacks.
Ingestion points: The data column from user-provided dataframes is ingested and processed in scripts/demo.py (e.g., in the LLM judge loop).
Boundary markers: The LLM judge prompt interpolates the untrusted content directly into the prompt without specific delimiters or instructions to ignore embedded instructions.
Capability inventory: The LLM output is used to determine relabeling actions within a local dataframe environment; there is no evidence of the data content being used to trigger shell commands or network exfiltration.
Sanitization: No input sanitization or validation is performed on the data content before it is passed to the LLM judge.
[EXTERNAL_DOWNLOADS]: The demonstration script fetches the ag_news dataset from Hugging Face's public repository, which is a well-known and established source for machine learning data.
[EXTERNAL_DOWNLOADS]: The skill relies on external libraries such as cleanlab, llm, and datasets, which are fetched from the official Python Package Index (PyPI).
[EXTERNAL_DOWNLOADS]: The assets/demo.html file loads frontend assets and JavaScript modules from the cdn.jsdelivr.net content delivery network to support the marimo notebook interface.

data-labeling-qa