Whitzard 白泽
Flagship Project

AgentGuard

Zero-trust access control for AI agent tool calls.

The Problem

AI agents with tool-use capabilities create runtime risks. When an LLM-powered agent can execute code, access files, or make API calls, a single prompt injection or misalignment can lead to unauthorized actions.

Traditional access control models assume trusted users. AI agents operate with delegated authority, making them a fundamentally different security challenge.

How It Works

01

Policy

Define fine-grained access control policies for every tool call an agent can make.

02

Trace

Full traceability of agent actions with structured audit logs.

03

Approval

Human-in-the-loop approval gates for high-risk operations.

04

Audit

Comprehensive audit trail for compliance and incident investigation.

Use Cases

Code Execution Agents

Control what code AI agents can execute and what resources they can access.

Browser Automation

Restrict browser agents from accessing unauthorized URLs or submitting forms.

API-Connected Agents

Enforce policies on which APIs agents can call and what data they can send.

Multi-Agent Systems

Coordinate access control across multiple collaborating AI agents.

Related: Nuwa Agent Safety Framework

AgentGuard implements runtime safety principles from the Nuwa Agent Safety Framework (NASF).

Learn more