nature-ref-verifier

Pass

Audited by Gen Agent Trust Hub on Aug 4, 2026

Risk Level: SAFEPROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from bibliographic files (BibTeX, RIS) and plain-text reference lists. Because this data is used to generate search queries and can trigger write operations to a Zotero library, it presents a surface where malicious metadata could attempt to influence the agent's logic.
  • Ingestion points: Bibliographic input files (BibTeX/RIS), Zotero item keys, and pasted reference lists as described in SKILL.md.
  • Boundary markers: Not explicitly implemented in the instructions; however, the skill follows a rigid multi-step verification workflow which provides some logical isolation.
  • Capability inventory: The skill can perform web searches (Bing/Google), query academic APIs (Crossref/IEEE), write to local files (Markdown/BibTeX patches), and perform CRUD operations on Zotero libraries via the pyzotero API or local HTTP API (SKILL.md).
  • Sanitization: No specific metadata sanitization or validation of the bibliographic fields is mentioned before they are processed by the agent.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 4, 2026, 02:28 AM
Security Audit — agent-trust-hub — nature-ref-verifier