Semantic Guardrails Architecture

Protegrity’s GenAI Security Semantic Guardrails solution is a security guardrail engine for AI systems. It evaluates risks in GenAI chatbots, workflows, and agents through advanced semantic analytics and intent classification to detect potentially malicious messages. PII detection can also be leveraged for comprehensive security coverage.

The documentation here for Semantic Guardrails covers its specific requirements and relationship with AI Developer Edition. For more information, refer to the complete body of the Semantic Guardrails documentation.

Overview

Semantic Guardrails is trained on synthetic customer-service AI chatbot datasets. The system performs best when analyzing conversations expected to match the training domain, that is, English-language-based customer service interactions involving orders, tickets, and purchases.

For domain-specific and user-specific applications requiring high detection accuracy, fine-tuning is necessary to completely leverage the model’s ability. This helps the model to learn from expected conversation patterns and message structures in both the inputs and outputs of protected GenAI systems.

The system operates by analyzing conversations between participants. These participants are users and AI systems, such as LLMs, agents, or contextual information sources. Furthermore, the system utilizes Protegrity’s Data Discovery, if present in the same network environment, to leverage PII detection in its internal decision algorithm.

The solution provides individual message risk scores and classifications, and cumulative conversation risk scores and classifications. This dual-scoring approach ensures that while individual messages may appear benign, potentially risky cumulative conversation patterns are identified. This significantly enhances detection of sophisticated attack vectors, including LLM jailbreaks and prompt injection attempts.

Architecture

For more information about the general architecture and working of Semantic Guardrails, refer to General architecture of Semantic Guardrails.