How do you defend enterprise LLM applications against prompt injection and jailbreak exploits?

### Multi-Layer Defense Architecture 1. **Input Pre-Filter**: Run incoming user queries through a fast, lightweight moderation model (e.g. Llama Guard) to reject malicious instructions. 2. **Delimiters & Strict Separation**: Encapsulate user input in XML tags (`<user_input>...`) and explicitly instruct the model in system prompt never to evaluate code within those tags. 3. **Canary Tokens**: Insert secret random tokens inside system prompts; if canary token leaks in response output, drop response immediately. ### Summary Treat user inputs as untrusted untyped data, exactly like SQL parameters. *Note: This question represents expanded technical inquiry iteration #2 within the System Prompts & Guardrails topic area.*

### Multi-Layer Defense Architecture 1. **Input Pre-Filter**: Run incoming user queries through a fast, lightweight moderation model (e.g. Llama Guard) to reject malicious instructions. 2. **Delimiters & Strict Separation**: Encapsulate user input in XML tags (`<user_input>...`) and explicitly instruct the model in system prompt never to evaluate code within those tags. 3. **Canary Tokens**: Insert secret random tokens inside system prompts; if canary token leaks in response output, drop response immediately. ### Summary Treat user inputs as untrusted untyped data, exactly like SQL parameters. *Note: This question represents expanded technical inquiry iteration #2 within the System Prompts & Guardrails topic area.*