Open Source · Apache 2.0

Open Source
AI Security

Code, models, tools, papers, and standards — all open under Apache 2.0. Open source discovers risk. The commercial runtime blocks it.

GitHubHugging Face

Open Source Tools

The Open Security Toolkit

Standalone tools for scanning, policy enforcement, evaluation, and integration.

og-scanScanner

Local scanner / CLI for skills, plugins, files, links, and risky actions. Outputs risk level, matched rules, explanations, and machine-readable JSON/SARIF.

Scan MCP tools and skills before installation
Check files and links for safety
Detect malicious patterns in agent inputs
Machine-readable output (JSON/SARIF)
og-policy-packPolicies

Default safety policies for different use cases. Define what your agent can and cannot do with pre-built and customizable policy sets.

Personal Safe Defaults
Solo Workflow Guard
Team Workspace Baseline
Enterprise Compliance Policies
og-evalsEvaluation

Real-world attack pattern benchmarks for testing your AI agent's security posture against known threats.

Malicious skill injection tests
Indirect prompt injection attacks
Data exfiltration scenarios
Tool misuse and abuse patterns
og-sdkSDK

Lightweight SDK for integrating OpenGuardrails protection into your AI agents and workflows.

scan_object() and check_action() APIs
Local confirmation hooks
Standardized trace output
MCP and OpenClaw integration

Platform & Models

Open Source Foundation

The core platform, guard models, research, and standards are all open.

GH

Platform Source

Full platform and deployment scripts on GitHub

GitHub →
HF

Guard Models

Guard models on Hugging Face under Apache 2.0

Models →
AP

Technical Paper

Full evaluation methodology and benchmark results

arXiv →
ST

AI-RSMS Standard

Community-driven runtime security standard

Standard →

Benchmarks

State-of-the-Art Performance

OpenGuardrails benchmark results

Average F1 scores across safety classification benchmarks. Full technical report →

87.1%
English Prompt F1
+2.8% vs next best
88.5%
English Response F1
+8.0% vs next best
97.3%
Multilingual Prompt F1
+12.3% vs next best
97.2%
Multilingual Response F1
+19.1% vs next best

Open Source. Enterprise Ready.

All open source components are released under the Apache 2.0 license.

Open source discovers risk. The commercial runtime blocks it.

Star on GitHub