The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill documentation encourages the use of trainer.load() and torch.load() for managing model checkpoints, as seen in references/training_evaluation.md. These functions rely on the pickle module for deserialization, which is inherently unsafe and can lead to arbitrary code execution if a malicious model file is loaded.
[INDIRECT_PROMPT_INJECTION]: The skill ingests large medical datasets which serve as an untrusted data surface. 1. Ingestion points: MIMIC4Dataset and related classes loading from local paths (e.g., SKILL.md line 125). 2. Boundary markers: No delimiters or explicit instructions to ignore embedded content were found in the dataset processing logic. 3. Capability inventory: The skill utilizes Bash(python:*) and Write capabilities. 4. Sanitization: No specific evidence of clinical text or metadata sanitization was observed before processing.
[COMMAND_EXECUTION]: The skill relies on Bash(python:*) for environment setup and model training, providing a direct execution path for potential payloads derived from other vulnerabilities.
[EXTERNAL_DOWNLOADS]: The skill installs pyhealth and its dependencies from public registries via uv pip install. While standard for development, this requires supply chain monitoring in sensitive healthcare environments.

alterlab-pyhealth