The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill's documentation and examples encourage the use of the pickle module for saving and loading analysis objects (DeseqDataSet). Python's pickle library is known to be insecure against erroneous or maliciously constructed data, potentially allowing for arbitrary code execution during deserialization.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-provided CSV, TSV, and AnnData files. This functionality represents an attack surface where malicious instructions could be embedded in experimental data or metadata to manipulate the agent's output or behavior.
Ingestion points: Files are read using pd.read_csv, pd.read_csv(sep="\t"), and ad.read_h5ad in the SKILL.md workflows.
Boundary markers: There are no instructions provided to the agent to treat data content as untrusted or to ignore instructions contained within data fields.
Capability inventory: The skill allows for file writing (to_csv, pickle.dump) and complex data transformations, which could be leveraged if an injection is successful.
Sanitization: No input validation or sanitization routines are implemented to check the integrity or safety of the data files before they are processed by the agent.

pydeseq2