The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill documents legitimate network operations for scientific purposes, such as fetching genomic sequences using genomepy or motif libraries. It also instructs the agent to fetch bibliographic metadata from arxiv.org to ensure accurate citations. These are standard operations within the scientific research context.
[DYNAMIC_EXECUTION]: The skill utilizes standard model persistence methods (scvi.model.SCVI.load). While the loading of serialized Python objects (often using pickle internally) is a potential vector for code execution, the skill explicitly includes a security warning to 'Load only trusted artifacts,' which aligns with industry best practices for machine learning workflows.
[DATA_EXPOSURE]: The skill instructions involve reading and writing local datasets (e.g., .h5ad files) and model checkpoints. These operations are restricted to the local file system and represent the primary functional purpose of the toolset.
[INDIRECT_PROMPT_INJECTION]: The skill defines an attack surface by ingesting external data files into the agent's context. However, it incorporates data validation steps (e.g., checking for unique identifiers and finite numerical values) and operates on structured scientific data rather than raw natural language prompts, minimizing the risk of indirect injection attacks.

scvi-tools