The Agent Skills Directory

[SAFE]: The skill provides legitimate technical documentation and automation scripts for a machine learning optimization workflow.
[COMMAND_EXECUTION]: The skill includes several Python scripts (compare_outputs.py, generate_qat_config.py, inspect_quant_checkpoint.py, inspect_quant_onnx.py) intended for local execution. These scripts process local files and perform validation tasks without making network connections or executing unauthorized shell commands.
[DYNAMIC_EXECUTION]: In inspect_quant_checkpoint.py, the skill utilizes torch.load to inspect model weights. It correctly implements security guardrails by using weights_only=True by default, which prevents arbitrary code execution through malicious pickle data. The script explicitly requires a user-provided flag to override this safety feature for older environments, accompanied by a clear warning.
[INDIRECT_PROMPT_INJECTION]: The skill defines workflows for ingesting and processing untrusted data artifacts like third-party model checkpoints. The risk is mitigated by explicit instructions to use trusted sources and the implementation of safe deserialization practices in the provided scripts (e.g., np.load(..., allow_pickle=False) in compare_outputs.py).

lnn-quant-training