The Agent Skills Directory

[METADATA_POISONING]: The skill package includes a file named 'pdb-database_audit_result_v1.json'. This file is a fabricated security report that uses simulated audit metrics (e.g., 'security: PASS', 'static_score: 84', 'deployable: true') to project a false sense of security and third-party verification. This deceptive practice is designed to misguide security reviewers and automated scanners.
[INDIRECT_PROMPT_INJECTION]: The skill ingests metadata and sequence data from the external RCSB PDB Data API, which constitutes an untrusted data surface.
Ingestion points: Data is retrieved via 'rcsbapi.data.fetch' and 'requests.get' in the example python script and documentation.
Boundary markers: Absent. There are no instructions to the agent to treat fetched metadata as untrusted or to ignore instructions embedded within entry titles or keywords.
Capability inventory: The skill uses 'pathlib' for file operations and interacts with the 'rcsb-api' client.
Sanitization: The provided implementation does not sanitize or validate strings retrieved from the API (such as 'struct.title') before processing them.
[EXTERNAL_DOWNLOADS]: The skill downloads protein structure files and metadata from RCSB PDB official domains (files.rcsb.org and data.rcsb.org). These are well-known and trusted scientific data repositories.

pdb-database