OpenClaw Needs Real Security Controls; We Built Them Open Source

AI agent adoption and development are evolving quickly. The tooling used to build agents is improving fast, but the security controls around those agents are often rigid, opaque, or difficult to adapt to real environments. As more teams experiment with OpenClaw, one challenge becomes clear: developers need ways to inspect what agents are doing, evaluate risky behavior, and intervene when necessary.

Today, Zenity released a new open-source security infrastructure for OpenClaw that enables teams to add detection and blocking capabilities directly into agent workflows.

This project comes out of Zenity’s ongoing research into how agents operate in real environments. While the Zenity platform provides enterprise-grade security and governance across AI systems, this open-source release focuses on a specific need emerging in the OpenClaw ecosystem, giving developers a practical way to instrument and evaluate agent behavior inside their own environments.

The result is a flexible framework that allows teams to plug in their own security logic, integrate existing tooling, and define what activity should be allowed, flagged, or blocked during agent execution.

A Security Framework You Can Extend

At its core, the framework provides a mechanism for evaluating activity inside an agent workflow and determining whether it should proceed.

These checks are implemented through evaluators, modular inspections that run during the agent execution process. An evaluator examines an event and determines whether the activity is acceptable from a security standpoint.

For example, an evaluator might scan inbound prompts for sensitive patterns such as social security numbers, credentials, or internal identifiers before that information ever reaches the agent. Another evaluator could inspect a command the agent is about to execute or analyze output generated by a downstream tool.

Instead of prescribing which checks should exist, the framework provides the infrastructure required to run those evaluations wherever they are needed.

Security Checks Across the Agent Execution Flow

Security evaluation becomes far more effective when it can observe activity throughout the agent lifecycle. The framework allows checks to run at several critical points in the execution flow.

One checkpoint occurs when messages enter the agent. Prompts and other inbound interactions can be inspected before they become part of the agent’s context. This allows teams to detect sensitive information or disallowed inputs before the agent begins reasoning about them.

Another checkpoint occurs before a tool is executed. Agents frequently interact with external systems through tool calls. Evaluating those calls before execution allows teams to identify potentially dangerous actions before they occur. A risky command, an unauthorized system interaction, or an operation that violates policy can be stopped before the agent carries it out.

A final checkpoint occurs after a tool executes, when the resulting output can be analyzed. Even when an action appears safe, the response may still expose secrets or sensitive information. Inspecting output provides an additional layer of control before the workflow continues.

Together, these inspection points allow organizations to evaluate what the agent receives, what it attempts to execute, and what it ultimately produces.

A Foundation for Safer Agent Workflows

OpenClaw is enabling developers to build increasingly capable agent-driven systems. As those capabilities expand, organizations need ways to observe agent behavior, apply security checks, and intervene when necessary.

This project provides a practical foundation for doing exactly that.

By enabling detection and blocking directly inside agent workflows, it allows teams to implement security controls where they matter most. The result is greater visibility, stronger safeguards, and more confidence when deploying agent-based systems.

Many of the ideas behind this work are also reflected in the Zenity AI Security Platform, where similar principles of visibility, detection, and control are applied to AI agents operating across SaaS, cloud, and endpoint environments.

The project is available now as open source. Explore the repository and see how detection and blocking can be integrated directly into your OpenClaw workflows.