How keremp/insecure-ai-agents works
Data Layer
• SQLite Database (Patient Data)
• Tables/fields:
• Patient Records (name, age, gender, diagnosis_date, etc.)
• External Data Source (Johns Hopkins Clinical Trials)
AI Agent Architecture
• Supervisor Agent
• Worker Agent 1 (Database Access)
• Worker Agent 2 (Web Scraping)
Integration Points
• Web Scraping Interface
• Database Interface
• Inter-agent Communication Channels
• [External Clinical Trials Data] → [Web Scraping Agent]
• ↓
• [Patient Database] → [Database Agent] → [Supervisor Agent] → [Results/Recommendations]
Security Boundaries (Critical for this project)
• Potential vulnerability points:
• Prompt injection entry points
• Agent communication boundaries
• Data access controls
Data Flow Representation
• Raw data flow (patient data, trial data)
• Agent communications
• Final recommendations flow
• Red: Security-critical components
• Blue: Data storage components
• Green: Processing agents
• Yellow: External data sources
Special Considerations:
Security Layer
• A security overlay showing:
• Prompt injection vulnerabilities
• Excessive agency risks
• Data access controls
Guardrails Implementation
• Where security measures are implemented
• Differences between the three versions:
• Basic version
• With guardrails
• With Haize
Process Flow
1. Database initialization
2. Trial data collection
3. Patient-trial matching
4. Recommendation generation
• Agent Boundaries
• Data Flow Types
• Security Zones
• Vulnerability Points
• Control Mechanisms
Key Architectural Patterns to Highlight:
Data Pipeline Pattern
• The high-level multi-agent architecture
• The security vulnerability points
• The data flow between components
• The security measures in different versions
Remember to include notes about:
• LangSmith integration points
• Security review findings
• Different versions of the implementation
This system design should emphasize both the functional architecture and the security considerations, as this is a demonstration project for AI security vulnerabilities.