The Agent Skills Directory

[EXTERNAL_DOWNLOADS] (SAFE): The documentation includes standard examples for reading data from remote URLs and S3 buckets using fsspec and urllib. These patterns are typical for data science workflows and do not execute remote code.- [INDIRECT_PROMPT_INJECTION] (LOW): The skill demonstrates extensive data ingestion capabilities from external files. 1. Ingestion points: 'references/io_operations.md' identifies multiple entry points (read_h5ad, read_csv, read_excel, etc.). 2. Boundary markers: No delimiters or warnings regarding embedded instructions are present in the examples. 3. Capability inventory: 'references/io_operations.md' includes file writing (write_h5ad, write_csvs) and network retrieval (urllib.request). 4. Sanitization: No sanitization or validation of the ingested data content is shown. This surface could be exploited if an agent processes files from untrusted sources that contain hidden instructions.

anndata