The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses joblib.load and torch.load to load model artifacts in both SKILL.md and the scripts/benchmark_model.py script. These methods utilize the pickle serialization format, which can be exploited to execute arbitrary code when loading files from untrusted sources.
[PROMPT_INJECTION]: The benchmarking script establishes an indirect prompt injection surface by ingesting and processing data from external CSV files.
Ingestion points: The scripts/benchmark_model.py script reads data files provided via the data command-line argument.
Boundary markers: There are no explicit delimiters or instructions provided to the agent or within the script to separate untrusted data from the processing logic.
Capability inventory: The skill possesses capabilities to execute shell commands (via Bash), read and write files, and perform network operations.
Sanitization: The input data is processed directly by the pandas library without additional validation or sanitization of the content.
[EXTERNAL_DOWNLOADS]: The skill instructions and guide references well-known machine learning libraries such as transformers and auto-gptq, which may fetch pre-trained models and configurations from official repositories like Hugging Face.

serve