NYC

nutritional-specialist

Pass

Audited by Gen Agent Trust Hub on Feb 15, 2026

Risk Level: LOWPROMPT_INJECTION
Full Analysis
  • Indirect Prompt Injection (LOW): The skill reads dietary and health preferences from a local JSON file and displays them to the agent, creating a surface for embedded instructions. 1. Ingestion points: Data is read from ~/.claude/nutritional_preferences.json via scripts/preferences_manager.py. 2. Boundary markers: Absent; user content is displayed in a plain text list format. 3. Capability inventory: Negligible; the main JavaScript entry point is a skeleton implementation. 4. Sanitization: None; input from the JSON file is used directly in the display output.
  • Data Exposure (INFO): The skill records PII including health conditions and allergies in the user's home directory. While appropriate for the skill's function, this storage should be considered sensitive.
Audit Metadata
Risk Level
LOW
Analyzed
Feb 15, 2026, 09:38 PM