AI Defense Mechanisms

AI Defense Mechanisms & Guardrails

PoC Architecture Design

This PoC details the implementation of robust defense architectures for LLM systems, leveraging Semantic Routers and Guardrails to mitigate adversarial attacks.

Core Mechanics

1. Semantic Routing: A fast vector search matches incoming prompts against a database of known malicious intents or topics. If a match is found, the request is blocked before reaching the main LLM.
2. Input/Output Guardrails (NeMo): A dedicated secondary LLM (e.g., Llama Guard) evaluates the safety of the user input and the generated output, enforcing strict conversational bounds.
3. Sanitization: Stripping control characters, limiting context length, and using delimiters (like XML tags <user_input>) to segregate instructions from data.

Architecture Map

%%{init: {"theme": "default", "flowchart": {"useMaxWidth": true}}}%%
flowchart TD
    subgraph InputPhase ["Input Processing"]
        A["Raw User Input"]
        B["Sanitization & Formatting (XML Tags)"]
    end
    
    subgraph SemanticRouter ["Fast Filtering"]
        C["Vector Embedding Model"]
        D["Cosine Similarity Check"]
        E["Block (Malicious Match)"]
    end
    
    subgraph Guardrails ["NeMo / Llama Guard"]
        F["Input Safety Evaluation"]
        G["Primary LLM Generation"]
        H["Output Safety Evaluation"]
    end
    
    A --> B
    B --> C
    C --> D
    D -->|"High Score"| E
    D -->|"Low Score"| F
    F -->|"Safe"| G
    F -->|"Unsafe"| E
    G --> H
    H -->|"Safe"| I["Return Response"]
    H -->|"Unsafe"| E