The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill utilizes library functions like ov.utils.download_GDSC_data() and ov.utils.download_CaDRReS_model() to fetch external datasets and pre-trained models required for drug response analysis.
[EXTERNAL_DOWNLOADS]: Instructions are provided to clone an external GitHub repository (https://github.com/CSB5/CaDRReS-Sc) to obtain supplementary scripts needed for the Drug_Response module.
[COMMAND_EXECUTION]: The ov.single.Drug_Response function takes a scriptpath argument pointing to the cloned CaDRReS-Sc directory, which indicates that it executes external code from that path.
[COMMAND_EXECUTION]: The skill uses ov.utils.save to persist analysis objects (e.g., scenic_obj), which likely employs Python's pickle module. This poses a risk of unsafe deserialization if these files are manipulated or retrieved from untrusted sources.
[REMOTE_CODE_EXECUTION]: The combination of cloning a remote repository and subsequently executing its contents via library functions constitutes a remote code execution vector from a non-whitelisted source.
[PROMPT_INJECTION]: The skill ingests untrusted biological data (AnnData, feather databases, motif tables) from external and local sources. This represents an attack surface for indirect prompt injection.
Ingestion points: AnnData objects via dataset loaders, *.feather databases, and *.tbl files in the SCENIC module.
Boundary markers: None detected in the prompt instructions to isolate data from agent instructions.
Capability inventory: Includes file writing (adata.write, ov.utils.save) and external code execution (ov.single.Drug_Response).
Sanitization: No specific sanitization or validation logic is applied to the content of ingested biological data before processing.

single-cell-downstream-analysis