The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external datasets (train, test, and query archives) provided as folder paths. These sources could theoretically contain malicious data patterns designed to influence agent behavior during training or evaluation tasks.
Ingestion points: dataset.train_dataset_dir, dataset.test_dataset_dir, and dataset.query_dataset_dir as defined in SKILL.md and references/skill_info.yaml.
Boundary markers: The instructions do not define specific delimiters or instructions to ignore embedded commands within these data inputs.
Capability inventory: The agent is granted access to the Bash tool to execute docker run commands for training and inference.
Sanitization: Input data is processed by the official re_identification CLI tool within the secured TAO Toolkit container.
[DYNAMIC_EXECUTION]: The skill documents the use of the environment variable TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1 to allow loading PyTorch checkpoints containing complex objects (OmegaConf containers). This allows for more flexible deserialization which could be exploited by malicious files.
Evidence: Noted in SKILL.md as a workaround for checkpoint loading failures.
Mitigation: The skill provides an explicit security warning advising users to "not use this env var for untrusted checkpoints," restricting its use to trusted, user-generated artifacts.
[COMMAND_EXECUTION]: The skill provides structured docker run commands to perform training and evaluation inside a containerized environment. These commands use standard NVIDIA TAO parameters for memory and GPU management.
[SAFE]: The skill relies on the official nvcr.io/nvidia/tao/tao-toolkit container image, which is a verified resource provided by the vendor (NVIDIA).
[SAFE]: The inclusion of an in-toto digital signature bundle (skill.oms.sig) demonstrates high supply-chain security maturity by allowing for content integrity verification.

tao-train-reid