The Agent Skills Directory

[SAFE]: The skill is designed to interact with official NVIDIA BioNeMo microservices. All network operations target legitimate endpoints such as health.api.nvidia.com and nvcr.io.
[DATA_EXPOSURE]: The skill manages credentials correctly by using environment variables (NGC_API_KEY) and suggesting the use of .env files for local development. It explicitly warns against hardcoding secrets and uses bearer authentication for hosted API calls.
[INDIRECT_PROMPT_INJECTION]: The skill handles user-supplied DNA sequences. The included script scripts/generate.py implements a clean_dna function that validates the sequence against the A/C/G/T alphabet and normalizes input, effectively sanitizing the data before it is sent to the inference endpoint.
[DYNAMIC_EXECUTION]: For processing tensor outputs from the local forward endpoint, the skill uses numpy.load with allow_pickle=False. This prevents unsafe deserialization of the binary data received from the NIM service.
[COMMAND_EXECUTION]: Local Docker deployment instructions use standard flags and environment variables. The use of chmod 700 on the local cache directory demonstrates a commitment to least-privilege security by restricting access to the owner.

evo2-nim