The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The documentation in references/hyperparameter-tuning.md includes a PostgreSQL connection string containing embedded credential placeholders: postgresql://user:pass@localhost/optuna. While intended as an example, this practice can lead to accidental exposure of sensitive information if the pattern is adapted for production use.
[REMOTE_CODE_EXECUTION]: Multiple documents, including SKILL.md and references/callbacks.md, demonstrate model checkpoint loading using the load_from_checkpoint method. This method internally utilizes PyTorch's torch.load, which defaults to Python's pickle module for deserialization. This is a known security risk as loading untrusted or maliciously crafted checkpoints can result in arbitrary code execution on the host machine.
[COMMAND_EXECUTION]: The hyperparameter tuning workflows described in references/hyperparameter-tuning.md utilize libraries such as Ray Tune and Optuna to execute arbitrary Python functions (e.g., train_fn, objective). This provides an interface for dynamic code execution within the environment.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by providing functionality to ingest and process external model checkpoint files. Ingestion points: LitModel.load_from_checkpoint referenced in SKILL.md and references/callbacks.md. Boundary markers: None identified; there are no explicit instructions or delimiters to distinguish between data and potentially embedded malicious instructions. Capability inventory: The skill uses L.Trainer which can invoke shell commands for distributed operations and performs file-system read/write tasks. Sanitization: No validation or integrity checks (like cryptographic signatures) are performed on the external checkpoint data before loading.

pytorch-lightning