alphafold-database

Warn

Audited by Gen Agent Trust Hub on Nov 4, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [COMMAND_EXECUTION]: The download_proteome function in SKILL.md constructs a shell command by interpolating the taxonomy_id variable into a string and executing it via subprocess.run(cmd, shell=True). This implementation is vulnerable to shell command injection if the taxonomy_id is provided by an untrusted source without sanitization.
  • [INDIRECT_PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection by processing external data and providing command execution capabilities.
  • Ingestion points: The skill retrieves metadata and identifiers from external APIs including the UniProt API (urllib.request.urlopen) and the AlphaFold REST API (requests.get) as documented in SKILL.md and references/api_reference.md.
  • Boundary markers: The provided instructions do not include delimiters or warnings to treat fetched biological data as untrusted or to ignore potential instructions embedded in protein metadata.
  • Capability inventory: The skill possesses capabilities for file system operations (open().write()), network communication (requests, urllib), and arbitrary shell command execution (subprocess.run).
  • Sanitization: There is no logic shown for validating or escaping input parameters (such as taxonomy_id, query, or alphafold_id) before they are incorporated into shell commands or API requests.
  • [EXTERNAL_DOWNLOADS]: The skill fetches protein structures, confidence metrics, and bulk datasets from the AlphaFold Database hosted by EMBL-EBI and Google Cloud Public Datasets. These are recognized, well-known services within the structural biology community.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Nov 4, 2026, 10:11 AM