OpenGuardrails is an open-source AI security gateway that prevents enterprise secrets from being unknowingly sent to external LLMs β without breaking how teams use AI.
It detects sensitive data, evaluates risk, and automatically blocks, anonymizes, or routes requests to private models β before they ever reach an external LLM.
Try instantly on our free SaaS, then deploy the same open-source code in your own production environment.
See how OpenGuardrails protects your AI applications in real-time
We're building the AI Runtime Security Management System (AI-RSMS) β an open, community-driven standard for securing AI systems during runtime. Join security, IT, risk, and compliance leaders worldwide to shape this critical standard together.
The open-source standard for securing enterprise AI inference at runtime
Built for real production environments β not demos.
"We've already crossed the gap from lab to production."
Fits naturally into modern open-source AI stacks.
"If your AI stack is open-source, your guardrails should be too."
Your policies. Your risks. Your guardrails.
"This is not content moderation β this is enterprise control."
OpenGuardrails is the open-source standard for securing enterprise AI inference at runtime.
Whether you're an enterprise AI developer, architect, or security lead β this is the only reasonable security layer choice.
Leading the industry in AI safety and guardrails performance
Dynamic per-request policy configuration with continuous probabilistic sensitivity thresholds (Ο β [0,1]). Unlike static "strict/loose" modes, OpenGuardrails enables real-time sensitivity tuning for safety governance.
Single 14Bβ3.3B (GPTQ quantized) model handling both content-safety and model-manipulation detection. Achieves superior semantic understanding compared to hybrid BERT-style architectures while maintaining production-level efficiency.
Robust performance across 119 languages and dialects, with SOTA results on English, Chinese, and multilingual benchmarks. Includes OpenGuardrailsMixZh 97k dataset contribution under Apache 2.0 license.
First fully open-source guardrail system with both large-scale safety LLM and deployable platform. RESTful APIs, Docker deployment, and modular components for seamless private/on-premise integration.
Everything you need to secure AI applications across any cloud or deployment
Protect AI models across AWS, Azure, GCP, and on-premise deployments. Works with OpenAI, Anthropic, open-source models, and custom LLMs - wherever they run.
RESTful API with SDKs for Python, Node.js, Java, and Go. Get started in minutes with comprehensive docs and code examples.
Advanced protection against jailbreaks, prompt injection, code-interpreter abuse, and malicious code generation attempts.
Detect harmful, hateful, illegal, or sexually explicit content across 12 risk categories with configurable sensitivity thresholds.
Identify and redact sensitive personal and organizational information using NER pipelines and regex-based detection.
P95 latency of 274.6ms ensures your applications stay fast. High concurrency support for production workloads.
Latest from the blog
Release notes, field insights, and security research from the OpenGuardrails team and partners.
Jan 23, 2026
Large Language Models have become the second brain of modern enterprises. But in real enterprise environments, one uncomfortable question keeps surfacing: do we actually know how much sensitive data is being sent to external LLMs β unintentionally?
Dec 27, 2025
A global call to shape AI Runtime Security together. OpenGuardrails announces the AI Runtime Security Management System (AI-RSMS) β an open, community-driven standard draft focused on securing AI systems during runtime.
Dec 16, 2025
OpenGuardrails 4.5.0 introduces Direct Model Accessβa privacy-first feature that lets enterprises quickly deploy private POCs by pointing to our SaaS models without logging any data. Deploy locally, access models remotely, keep everything private.
