The Agent Skills Directory

[METADATA_POISONING]: The metadata in SKILL.md lists the author as 'K-Dense Inc.', which does not match the system author context 'eturkes'. This inconsistency may lead to confusion regarding the skill's origin.
[INDIRECT_PROMPT_INJECTION]: The skill's architecture is built around processing external data, which introduces a surface for indirect prompt injection attacks.
Ingestion points: scripts/template_datamodule.py and references/data_module.md provide templates for loading data from local paths and remote URLs.
Capability inventory: The skill is configured with the Bash tool, meaning that if a malicious instruction were successfully injected via processed data, it could potentially lead to command execution.
Boundary markers: There are no specific delimiters or instructions provided to the agent to treat external data as untrusted or to ignore instructions contained within it.
Sanitization: The provided code templates do not include validation or sanitization logic for ingested data.
[DYNAMIC_EXECUTION]: The skill instructions (e.g., in references/lightning_module.md) describe the use of load_from_checkpoint. This function typically uses torch.load, which is based on the pickle module and is vulnerable to arbitrary code execution if used to load untrusted model files.

pytorch-lightning