The Agent Skills Directory

EXTERNAL_DOWNLOADS (LOW): The skill performs network operations using requests.get and the rcsb-api package to fetch data from rcsb.org and its subdomains (files.rcsb.org, data.rcsb.org). While these are reputable scientific sources, they are not on the pre-defined trust whitelist.
COMMAND_EXECUTION (MEDIUM): The skill demonstrates the ability to write downloaded content directly to the local file system (e.g., open(f"{pdb_id}.pdb", "w").write(response.text)). This capability, combined with external data ingestion, presents a risk if the source were compromised.
INDIRECT_PROMPT_INJECTION (MEDIUM): In line with Category 8 analysis:
Ingestion points: Data enters the agent context via fetch() (GraphQL/REST) and requests.get() from external PDB servers.
Boundary markers: None present. The retrieved metadata (titles, descriptions, organism names) is processed and printed without delimiters.
Capability inventory: The skill possesses network read capabilities and local file-write capabilities.
Sanitization: No sanitization or validation of the retrieved text is observed before it is used in logic or displayed. A malicious entry in the PDB database could theoretically influence the agent's behavior.
UNVERIFIABLE_DEPENDENCIES (LOW): The skill requires rcsb-api, rcsbsearchapi, requests, and biopython. These are standard tools in the bioinformatics community but should be pinned to specific versions to prevent supply chain attacks.

pdb-database