The Agent Skills Directory

[PROMPT_INJECTION]: The skill instructions do not contain any attempts to override agent behavior, bypass safety guidelines, or extract system prompts.
[CREDENTIALS_UNSAFE]: No hardcoded secrets, API keys, or credentials were found. The encryption_key property in the schemas and spec templates is a configuration placeholder for user-supplied encryption keys.
[DATA_EXFILTRATION]: No unauthorized network operations or data transmission patterns were detected. Examples use placeholder S3 paths for datasets, which is consistent with the skill's purpose for model training.
[REMOTE_CODE_EXECUTION]: The skill does not download or execute remote scripts. It utilizes a static container image (tao_toolkit.pyt) for its execution environment.
[COMMAND_EXECUTION]: Shell commands are appropriately scoped to the TAO Toolkit's pose classification actions (train, evaluate, export, inference) as defined in the skill metadata.
[SAFE]: Evaluation of potential indirect prompt injection surfaces identifies it as standard data ingestion for machine learning training.
Ingestion points: Dataset data and label paths defined in SKILL.md and references/skill_info.yaml.
Boundary markers: Not present, as the input data is expected in specific binary or serialized formats (.npy, .pkl).
Capability inventory: Execution of defined TAO CLI tools inside a managed container.
Sanitization: The TAO Toolkit processes structured skeleton data, which limits the risk of instruction injection via input files.

tao-train-pose-classification