OpenGuardrails

OpenGuardrails (OGR) is an open, vendor-neutral protocol for AI agent safety and security, together with a neutral benchmark for comparing guardrail vendors. It is not itself a guardrail product. It defines a common contract that lets organisations integrate safety and security checks once and enforce them across every AI agent and large language model (LLM) they run. Without it, each agent, detection vendor and LLM protocol would have to be connected to every other by hand. The protocol covers two domains under one contract. Safety covers harmful content and behaviour, such as toxicity or self-harm. Security covers system compromise, such as prompt injection, data exfiltration, malicious commands and secret leakage.

The protocol models agent traffic as six layers, much as network firewalls model packets: session, turn, step, event, tool call and real-world execution. For each model call, an integration sends two “events” to a runtime endpoint. The first is the request, before the model sees it. The second is the response, before the agent acts on it. Each event carries the raw provider message and an identity tuple naming the agent, its type, its workspace and its user. The runtime returns a verdict, such as allow or block, at a point where the action can still be refused. Sessions and turns are reconstructed on the server.

Detector plugins from different vendors return verdicts that are combined under defined composition rules, and risks are labelled using a versioned taxonomy. Further specifications cover what agents may do (mandates), what they may claim (grounding), and behaviour when the runtime is unreachable, which by default lets the agent continue. Reference integrations exist for gateways, bridges and agent frameworks. A benchmark ranks conformant detectors on shared datasets.

Auto-discovered on 2026-10-14 by OECD Catalogue Automation