The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data that is interpolated into the model's context, creating a potential surface for indirect prompt injection.
Ingestion points: Metadata JSON files (train.json, val.json) and text files (prompts.txt) as described in SKILL.md and referenced in YAML spec templates.
Boundary markers: Absent. The instructions do not prescribe the use of delimiters or warnings to ignore instructions embedded within the metadata or prompt files.
Capability inventory: The skill possesses significant capabilities, including docker run execution with host network and GPU access, and the ability to write result files to the local file system.
Sanitization: No explicit sanitization or validation of the content within the JSON caption fields or text prompt files is mentioned.
[COMMAND_EXECUTION]: The skill utilizes high-privilege shell commands to manage the AI training environment.
Evidence: SKILL.md defines DOCKER_COMMON flags including --gpus all, --ipc=host, and --network=host, which grant the container extensive access to host resources. These are standard for NVIDIA TAO high-performance training but represent a high-capability execution environment.
Evidence: The environment variable TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1 is explicitly set to bypass the default security behavior in PyTorch 2.6+ that restricts loading non-safe pickle objects, which is required for compatibility with certain TAO checkpoints.
[EXTERNAL_DOWNLOADS]: The skill and its evaluation configuration download resources from established vendor and service domains.
Evidence: Fetches container images from NVIDIA's official registry at nvcr.io.
Evidence: The evaluation config (eval.config) uses s5cmd to mirror datasets and models from an internal NVIDIA S3 endpoint (pdx.s8k.io).

tao-finetune-video-clip