The Agent Skills Directory

[SAFE]: The skill references and downloads ML model checkpoints and configuration files from official NVIDIA organizations on Hugging Face (e.g., nvidia/NV-CodonFM-Encodon-80M-v1). These are well-known, trusted sources associated with the skill's author, and the downloads are documented neutrally for their intended purpose.
[INDIRECT_PROMPT_INJECTION]: The skill processes external CSV files provided by the user. It includes specific instructions to validate headers and verify that reference sequences match the expected codons at the specified positions before execution. This practice reduces the risk of the agent acting on malformed or malicious data embedded in the sequences.
[COMMAND_EXECUTION]: The skill prepares shell commands to run a local ML inference runner (src.runner). The instructions emphasize using existing Python environments and standard library tools (zipfile, tempfile) rather than assuming the presence of arbitrary system binaries, adhering to the principle of least privilege.
[EXTERNAL_DOWNLOADS]: The skill fetches model weights and configuration files from Hugging Face. As these resources belong to the verified vendor (NVIDIA) and are necessary for the skill's primary function of variant scoring, they are considered safe.

codonfm-score