The Agent Skills Directory

[DATA_EXFILTRATION] (MEDIUM): Potential Path Traversal in file writing.
Evidence: In scripts/gene_analysis.py, output_prefix is derived directly from gene_name (line 19: output_prefix = gene_name.lower()) and used to construct file paths (line 42: f"{output_prefix}_info.csv"). An attacker providing a gene name like ../../tmp/malicious could write files to arbitrary locations.
Evidence: In scripts/enrichment_pipeline.py, the output_prefix argument (line 46) is used directly in file paths (line 72: f"{output_prefix}_{db_key}.csv") without validation or sanitization.
[PROMPT_INJECTION] (MEDIUM): Indirect Prompt Injection surface (Category 8).
Ingestion points: The scripts fetch data from several external, third-party controlled sources including Enrichr, NCBI (BLAST), ARCHS4, and OpenTargets (found in scripts/gene_analysis.py and scripts/enrichment_pipeline.py).
Boundary markers: None identified. Data from these APIs is written directly to CSV/FASTA files and printed to the console.
Capability inventory: The scripts have file system write access (Path.mkdir, open(..., 'w'), to_csv) and perform network operations via the gget library.
Sanitization: Absent. External content is interpolated into output files and logs without escaping or filtering.
Risk: If an AI agent reads the output of these scripts to provide summaries or make decisions, a malicious record in a public database (e.g., a poisoned gene description in Ensembl or a BLAST hit) could influence the agent's behavior.

gget