The Agent Skills Directory

[DYNAMIC_EXECUTION]: The perform_forecasting function and fine-tuning scripts load data normalization statistics from .pkl files (e.g., standardizer.pkl). Python's pickle module, typically used to handle these files, can execute arbitrary code during deserialization. While the skill defaults to downloading these from a trusted repository on Hugging Face, it also supports loading local files (e.g., from artifacts/) that could be untrusted.
[INDIRECT_PROMPT_INJECTION]: The skill provides an attack surface for indirect prompt injection by processing untrusted time-series data from CSV files.
Ingestion points: Untrusted data enters via pd.read_csv and CLI flags like --csv or --train-csv across multiple files including SKILL.md and references/automl.md.
Boundary markers: There are no documented delimiters or instructions for the agent to ignore potentially malicious content embedded within the data files.
Capability inventory: The skill environment allows for shell command execution (via uv run), file writing to the local system, and network access to Hugging Face.
Sanitization: There is no explicit evidence of data validation or sanitization before the CSV content is processed by the model or used in downstream scripts.
[EXTERNAL_DOWNLOADS]: The skill automatically fetches model checkpoints (.pt) and standardizers (.pkl) from the official nvidia/nv-tesseract-forecasting repository on Hugging Face. This operation targets a well-known service and a trusted organization, representing standard functionality for obtaining pretrained model weights.

tao-finetune-nv-tesseract-forecasting