OpenGuardrails is the trust layer for AI agents — protecting every action your agent takes in the real world.
Scan untrusted objects. Check risky actions. Require approval before damage happens. From personal agents to business-critical workflows.
Proven Performance
OpenGuardrails achieves SOTA results across multilingual safety benchmarks, outperforming LlamaGuard, Qwen3Guard, and other leading guard models.
Average F1 scores across safety classification benchmarks. Full technical report →
Single 14B dense model quantized to 3.3B via GPTQ. Handles both content-safety and manipulation detection with superior semantic understanding.
Dynamic per-request policy with continuous sensitivity thresholds. Tune precision-recall trade-offs in real time via probabilistic logit-space control.
Robust multilingual coverage with SOTA results on English, Chinese, and cross-lingual benchmarks. Includes 97k Chinese safety dataset contribution.
P95 latency of 274.6ms with high concurrency. GPTQ quantization enables real-time inference at enterprise scale without sacrificing accuracy.
Blog
Release notes, security research, and insights on securing AI agents in production.
Mar 11, 2026
The era of buying security platforms from vendors is ending. The era of building your own — with AI — has begun. Today we are releasing OpenKai, an open-source project that transforms autonomous agent runtimes into cybersecurity-focused agentic AI platforms.
Mar 10, 2026
An analysis of ClawHub's top Skills reveals a fundamental shift: AI Agents are evolving from chat interfaces into full-fledged operating systems capable of connecting software, executing tasks, and continuously self-improving.
Mar 5, 2026
Today we're releasing AI Agent Discovery, a new open-source project that helps organizations discover and inventory all AI agents running within their enterprise environment by integrating with existing EDR infrastructure.
Free helps you try OpenGuardrails. Personal helps you rely on it. Solo helps you run your business on it.