fact-checking

Pass

Audited by Gen Agent Trust Hub on Jul 21, 2026

Risk Level: SAFEPROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it processes untrusted documents and passes their content to sub-agents.
  • Ingestion points: Research paper files are read from the local file system using the Read tool in Phase 1.
  • Boundary markers: Extracted claims are provided to sub-agents in JSON format without explicit delimiters (e.g., XML tags) or specific warnings to disregard embedded instructions within the document text.
  • Capability inventory: The skill uses Agent for sub-tasking, WebSearch and WebFetch for internet data retrieval, and Write for file creation.
  • Sanitization: The skill does not sanitize, filter, or escape external content from the documents before its inclusion in sub-agent prompts.
  • [EXTERNAL_DOWNLOADS]: The skill utilizes the WebSearch and WebFetch tools to retrieve information from external web sources. This behavior is consistent with the skill's stated purpose of verifying claims against online evidence using the platform's native tools.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 21, 2026, 09:31 AM
Security Audit — agent-trust-hub — fact-checking