The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute docker run commands with the --network=host flag. While this is standard for certain TAO performance requirements, it bypasses network isolation between the container and the host.
[EXTERNAL_DOWNLOADS]: The skill pulls the nvcr.io/nvidia/tao/tao-toolkit container image. As this is an official resource from the vendor (NVIDIA), this finding is considered safe and part of the skill's intended functionality.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied Parquet files, creating a surface where malicious instructions could be embedded in data.
Ingestion points: The contents of the Parquet file pointed to by input_parquet in the experiment specification.
Boundary markers: None present; the skill does not include instructions to the model to ignore data-embedded directives.
Capability inventory: The skill allows the agent to execute shell commands (Bash tool) and run local Python scripts.
Sanitization: The verify_image_embeddings_spec.py script validates the YAML structure and checks for path existence but does not inspect or sanitize the data records within the Parquet files.

tao-generate-image-embeddings