Home / Platform / Soteria Guard
Runtime Protection

Soteria Guard: the AI firewall for every prompt and response

Guard runs inline and inspects traffic between users, applications and models. It stops attacks and data leaks before they reach your model or your customers.

Why you need an AI firewall

Traditional WAFs look for SQL injection and malicious payloads. LLM attacks are written in plain language, so they hide inside documents, emails, web pages and tool outputs. Guard understands the semantics of each request and of each response.

Core capabilities

Prompt Injection Defense

Multi-layer detection of direct injection, indirect injection inside retrieved content, and multi-turn jailbreaks.

Sensitive Data Protection

Detects and redacts PII, PHI, PCI, credentials, API keys and source code in both directions.

Content & Brand Safety

Toxicity, bias, self-harm, off-topic and competitor-mention policies, configurable per application.

Grounding & Hallucination Checks

Verifies RAG answers against source documents and flags unsupported claims before delivery.

How it is deployed

  • SDK for Python, Node.js, Java and Go, which takes a few lines of code
  • REST API for any language or low-code platform
  • Reverse proxy in front of OpenAI-compatible endpoints, with no code changes
  • Gateway plugin for Kong, Envoy, NGINX, LiteLLM and cloud API gateways
Guard is designed for a low-latency inline path. Most checks finish in tens of milliseconds, and streaming responses are inspected chunk by chunk.

Policy as code

Define policies in YAML or in the console. Version them in Git, roll them out per environment and audit every decision. Each policy can block, mask, warn or log.

Related

Securing a customer chatbot?

See how Guard fits into a complete GenAI application security program.

GenAI Applications →

Need help integrating?

Our FDE Pod can deploy and tune Guard inside your architecture.

FDE Pod →