GLC PromptGuard: Comprehensive Guide

Overview

PromptGuard is an eight-layer security gate designed to check user prompts, RAG chunks, and tool outputs before they reach the main model or agent tool loop. It operates as a critical security layer, evaluating intent, source context, and potential impact rather than relying solely on keyword detection. This tool helps protect systems from prompt injection attacks by analyzing untrusted text slices before they are processed.

Key Characteristics:-Service ID:promptguard-Version:0.3.25-Status:Available-Endpoint:https://mcp.glc-rag.hu/mcp (streamable HTTP)

Benefits

PromptGuard offers several key advantages for securing AI systems. It evaluates untrusted text slices only, not full chat history, which makes it efficient and focused. The system checks intent, source context, and potential impact to provide a security score and intent analysis. It supports multilingual detection, meaning it can analyze text in various languages without switching models or rules. The tool provides clear feedback on whether an injection attempt has been detected, along with details about the policy applied and any specific areas of concern. This helps developers and users understand the nature of potential threats before they reach the main model.

Use Cases

PromptGuard is designed for use in environments where AI models interact with untrusted inputs. It should be called from an orchestrator before every main LLM call to ensure security. The tool is particularly useful for checking user prompts, RAG chunks, and tool results. It can be integrated into MCP setups by adding the appropriate configuration to the mcp.json file. Developers can use the promptguard_check tool to evaluate text for prompt injection attacks. The system is also suitable for environments that require multilingual support, as it does not rely on language-specific models for detection. Users can register as agents for auto-approved access or as human users who require system-admin approval.

Pricing

The pricing model for PromptGuard involves a credit system. The base cost is one credit plus any LLM usage. In degraded mode, where the classifier or canary is skipped, the cost is still one credit if no LLM is used. This ensures that even in fallback scenarios, the system maintains a consistent billing structure. Users should be aware that sending full message arrays or asking the main LLM to call the tool first is not recommended and may lead to unexpected results.

Vibes

While there are no specific reviews or testimonials available in the provided information, the tool has been designed with a focus on comprehensive security and ease of integration. The developers have emphasized the importance of combining PromptGuard with deterministic tool-policy engines for complete protection. The system includes features like hard blocks for untrusted text flagged as exfiltration, financial, or destructive, which adds an extra layer of security. The availability of detailed guides and documentation suggests a commitment to user support and transparency.

Additional Information

PromptGuard is part of the GLC RAG ecosystem and is available through the Model Context Protocol. The service is currently in version 0.3.25 and is marked as available. Users can register as agents for auto-approved access or as human users who require system-admin approval. The tool includes features like spotlight.facts for RAG chunks and tool results, which helps in managing sensitive information. The system also includes error handling for various conditions, such as empty text, invalid context, and authentication failures. For more detailed information on the technical methodology, users can refer to the provided guides and documentation.