pubmed-database

Warn

Audited by Gen Agent Trust Hub on Mar 10, 2026

Risk Level: MEDIUMPROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [PROMPT_INJECTION]: Metadata Poisoning
  • The 'skill-author' field in the metadata is set to 'K-Dense Inc.', which contradicts the recognized author context of 'sickn33'.
  • [PROMPT_INJECTION]: Obfuscation
  • The technical support email address is hidden using a reversed string: 'vog.hin.mln.ibcn@seitilitue' (decodes to 'eutilities@ncbi.nlm.nih.gov').
  • [PROMPT_INJECTION]: Indirect Prompt Injection Surface
  • Ingestion points: The skill retrieves article IDs and abstracts from the external PubMed database (Capability 4).
  • Boundary markers: There are no explicit delimiters or instructions to ignore embedded content when processing retrieved abstracts.
  • Capability inventory: The skill is designed for network operations via the 'requests' library.
  • Sanitization: No sanitization, validation, or filtering of the retrieved third-party content is documented or implemented in the provided logic.
  • [EXTERNAL_DOWNLOADS]: Well-Known Service Reference
  • The skill programmatically fetches biomedical data from the official NCBI E-utilities API at 'ncbi.nlm.nih.gov'.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Mar 10, 2026, 08:24 AM