The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The documentation in references/r_interop.md contains instructions to use sudo for installing system dependencies on Linux (e.g., sudo apt-get install, sudo dnf install). While these are intended for environment setup, they represent a path to privilege escalation if executed by an agent.\n- [REMOTE_CODE_EXECUTION]: The skill instructs the agent to install R packages directly from a GitHub repository using remotes::install_github("mojaveazure/seurat-disk") in references/r_interop.md. This facilitates the download and execution of code from an external source not identified as a trusted organization.\n- [EXTERNAL_DOWNLOADS]: The skill makes extensive use of external package registries for its setup, including PyPI (scanpy, harmonypy, bbknn), CRAN (BiocManager, remotes), and Bioconductor (SingleCellExperiment, zellkonverter).\n- [COMMAND_EXECUTION]: The toolkit includes scripts that wrap and execute shell commands, specifically for R interoperability and system package management tasks.\n- [INDIRECT_PROMPT_INJECTION]: The skill has a significant attack surface for indirect prompt injection as it processes complex, researcher-provided biological data files (H5AD, RDS, CSV).\n
Ingestion points: Multiple scripts (scripts/convert.py, scripts/run_pipeline.py, and scripts/qc_analysis.py) load data from external files into the AnnData structure.\n
Boundary markers: No specific boundary markers or "ignore instructions" warnings are implemented for the data ingestion phase.\n
Capability inventory: The skill can write files to the local system, create directories, and execute external scripts (R code) which could be influenced by input metadata.\n
Sanitization: There is no evidence of sanitization for strings or metadata labels extracted from input datasets, which are often used in downstream plotting and annotation tasks.

scanpy