Agent authorization infrastructure

AI agents should not get unlimited authority.

NEXO Privilege Gate is an open-source project exploring independent authorization for AI agents. This public preview evaluates proposed actions, applies least-privilege policies, explains risk, and exposes when a human should remain in control.

Try policy preview View GitHub ↗

⛨ Least privilege

Authority should be granted per action, not assumed globally.

◎ Human control

Consequential actions should pause for explicit approval.

◇ Explainability

Decisions expose the policy and reason behind them.

→ Local-first roadmap

v0.4 is adding real filesystem enforcement inside an isolated sandbox.

Agent action request

Describe an action an AI agent wants to perform.

Authorization decision

Browser policy preview.

Decision

WAITING

Submit an action to run the policy preview.

Risk score 0 / 100

Decision history

Stored only in memory for this browser session.

No actions evaluated yet.

Security model

The project is evolving from policy simulation toward real enforcement.

Independent authorization

The long-term design separates the component requesting an action from the component deciding whether that action may execute.

Consequential-action review

Destructive, privileged, financial and sensitive-data actions are candidates for explicit human approval.

Verifiable history

The local v0.4 branch is introducing cryptographically chained audit records and real sandbox filesystem controls.

v0.4 architecture

Under active development in the public repository.

AI Agent

→

NEXO Gate

→

Policy Engine

→

Human Approval

→

Protected Tool

Current public page: browser-based policy simulation.

v0.4 local target: real filesystem operations inside an isolated sandbox, path-traversal protection, human approval for state-changing operations, quarantine instead of permanent deletion, and a tamper-evident local audit chain.

The project does not currently claim to protect arbitrary Windows, browser, network or MCP actions.