The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the user to clone a third-party GitHub repository and run a training script locally. This is identified as a remote code execution vector.
Evidence: git clone https://github.com/jquesnelle/yarn followed by python scripts/train.py in SKILL.md and references/fine_tuning.md.
Context: The repository belongs to a researcher cited in the skill's technical documentation. However, it is not on the pre-approved trusted vendor list, and the instructions promote executing the code with local environment permissions.
[INDIRECT_PROMPT_INJECTION]: The skill defines procedures for fine-tuning models on large, untrusted external datasets, which creates a significant attack surface for indirect prompt injection.
Ingestion points: references/fine_tuning.md describes loading data via load_dataset (e.g., pg19, arxiv, codeparrot/github-code).
Boundary markers: No explicit boundary markers or instructions to ignore embedded commands are present in the training data preparation snippets.
Capability inventory: The skill utilizes model training (trainer.train()) and text generation (model.generate()) capabilities.
Sanitization: There is no evidence of content sanitization or filtering applied to external dataset text before it is tokenized and processed by the model.
[EXTERNAL_DOWNLOADS]: The skill downloads and installs several Python packages from public registries.
Evidence: pip install transformers torch, pip install einops, pip install rotary-embedding-torch, and pip install flash-attn --no-build-isolation.
Context: While transformers and torch originate from trusted organizations, other packages such as einops and rotary-embedding-torch are maintained by third-party individuals and are installed at runtime.

long-context