This repository demonstrates basic AI safety controls for backend applications that integrate with LLMs.

• Prompt injection detection
• Sensitive data redaction
• Grounded prompt construction
• Citation requirement checks
• Output leakage checks
• Audit events for generation and blocked requests

• ignore previous instructions
• reveal system prompt
• bypass security
• exfiltrate secrets
• disable validation
• print environment variables

The current detector is rule-based and intentionally simple. It is useful for demonstrating the security boundary and testing strategy, but it is not sufficient for production by itself.

A production system should add:

• model-assisted classification
• allow/deny tool policies
• structured output schemas
• tenant isolation
• rate limiting
• authentication and authorization
• secrets management with Azure Key Vault or equivalent
• monitoring and alerting for suspicious prompts