The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructions and action configurations include a command to dynamically install a specific version of a library (pip install "protobuf<7") within the container environment to resolve a version incompatibility between components.
[DYNAMIC_EXECUTION]: A troubleshooting script in the documentation uses torch.load() with weights_only=False to verify adapter training status on a locally generated checkpoint file. This performs unsafe deserialization of pickle data, which is a standard but technically insecure practice in ML development.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied dataset metadata files (e.g., msrvtt_test_1k.json, labels.json) which contain free-form text captions. This represents a surface where malicious instructions in the training data could influence agent behavior.
Ingestion points: Metadata files are read as inputs for training, evaluation, and inference via the cosmos-embed1 CLI.
Boundary markers: No specific delimiters or instructions to ignore embedded commands are used when processing the caption text.
Capability inventory: The skill utilizes the cosmos-embed1 CLI tool and has the ability to perform dynamic package installations.
Sanitization: The documentation does not describe any sanitization or validation of the text content within the ingested metadata files.
[EXTERNAL_DOWNLOADS]: The skill pulls a container image from NVIDIA's NGC registry and downloads model weights from Hugging Face. Both are well-known and vendor-controlled sources.

tao-finetune-cosmos-embed