The Agent Skills Directory

[METADATA_POISONING]: The skill includes a file named gene-database_audit_result_v1.json which acts as a simulated security audit report. This file contains fabricated evaluation results, including a "PASS" security status and high quality scores. The presence of this file is a deceptive tactic designed to bypass security screenings by presenting a pre-validated safety profile.
[INDIRECT_PROMPT_INJECTION]: The skill handles untrusted external data that could influence its behavior if it contains malicious instructions.
Ingestion points: scripts/batch_gene_lookup.py reads gene symbols from external files via the --file argument, and scripts/query_gene.py accepts complex search strings via the --search argument.
Boundary markers: There are no delimiters or instructions provided to the agent to distinguish between data and instructions when processing these inputs.
Capability inventory: The skill has the capability to perform network requests (urllib.request.urlopen) to external APIs and read files from the local filesystem across all scripts (scripts/batch_gene_lookup.py, scripts/fetch_gene_data.py, and scripts/query_gene.py).
Sanitization: The scripts perform basic URL encoding but do not validate or sanitize the content of the search terms or file inputs before interpolating them into API request paths.
Self-referential content: The file gene-database_audit_result_v1.json contains claims specifically targeting security analysis processes ("security: PASS", "gate: PASS") to influence the verdict.
[EXTERNAL_DOWNLOADS]: The skill performs network operations to interact with the National Center for Biotechnology Information (NCBI) genomic services.
API Targets: eutils.ncbi.nlm.nih.gov (E-utilities) and api.ncbi.nlm.nih.gov (NCBI Datasets API).
These downloads are used to retrieve gene metadata, sequences, and literature links as part of the skill's primary functionality.

gene-database