The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides standardized templates and guides for building data pipelines (e.g., scripts/template_datamodule.py) that ingest external datasets for model training and validation. This creates a surface for indirect prompt injection, as malicious instructions embedded within the processed data could attempt to influence the agent's behavior during the training or evaluation process.
Ingestion points: Data processing methods such as prepare_data and setup in LightningDataModule, along with batch processing hooks in LightningModule (training_step, validation_step, test_step).
Boundary markers: The provided code templates do not include specific delimiters or instructions to the model to ignore embedded commands or distinguish between data and instructions.
Capability inventory: The skill possesses the capability to write to the local file system (via ModelCheckpoint), communicate with external logging services (Weights & Biases, MLflow, Comet), and execute shell commands (as the Bash tool is allowed in the skill configuration).
Sanitization: The provided instructions and boilerplate scripts do not implement explicit sanitization, filtering, or validation of the input data to mitigate potential injection attacks.

pytorch-lightning