The Agent Skills Directory

[SAFE]: No security issues were detected during the analysis of this skill. The skill provides instructions and schemas for training, evaluating, and running inference on weakly-supervised segmentation models using NVIDIA's TAO Toolkit.
[PROMPT_INJECTION]: The skill instructions do not contain any patterns attempting to bypass safety filters, override agent behavior, or extract system prompts. All instructions are focused on the intended machine learning workflow.
[CREDENTIALS_UNSAFE]: No hardcoded secrets, API keys, or private keys were found. Configuration templates and schemas correctly use empty strings or placeholders for sensitive fields like encryption_key.
[REMOTE_CODE_EXECUTION]: The skill utilizes established TAO Toolkit CLI tools (e.g., mal train, mal evaluate) for its operations as defined in references/skill_info.yaml. No unauthorized remote script execution or unverifiable package installations were detected.
[INDIRECT_PROMPT_INJECTION]: The skill processes image datasets and COCO-style annotation files. While this represents a standard data ingestion surface for machine learning tasks, no specific vulnerabilities were identified.
Ingestion points: Dataset and annotation paths (e.g., dataset.train_ann_path, inference.ann_path) defined in references/skill_info.yaml and processed during training or inference actions.
Boundary markers: The skill uses standard configuration parameters for data paths without specific prompt delimiters for the data content itself.
Capability inventory: Capabilities are limited to executing the mal toolkit commands within the environment's shell, as specified in the skill's action policy.
Sanitization: The skill relies on the underlying TAO Toolkit's internal validation for dataset integrity and formatting.

tao-train-mask-auto-label