Security

ZeroSphere is designed with a security-first architecture. This page explains how the system is built to protect you.

Visual-Only Agent Interaction

When the agent operates your application inside a virtual display, it sees the screen exactly the way a human does - raw pixels captured natively by the system. Nothing more.

The agent does not use:

  • Accessibility APIs (AT-SPI, UIAutomation, etc.)
  • The DOM or browser internals
  • Element IDs, CSS selectors, or JavaScript injection
  • Any hook into the application's memory or process

This means the agent cannot be given hidden instructions by a malicious web page, a crafted document, or a compromised application. It sees what is visually rendered on screen - nothing that is invisible to you is visible to it.

Prompt Injection Immunity

Prompt injection attacks attempt to embed hidden instructions inside content the AI reads - typically through invisible text in web pages, documents, or API responses.

ZeroSphere's virtual display eliminates this attack surface entirely:

  • The agent reads the screen as an image, the same way a human does
  • Text that is not rendered visibly on screen is not seen by the agent
  • There is no DOM parsing, no text extraction, no hidden layer for an attacker to exploit

A malicious web page cannot instruct the agent to do anything by hiding text in the HTML. If it is not visible on screen, it does not exist to the agent.

You Control What the Agent Launches

The agent cannot open applications or start virtual sessions without your permission. All virtual session creation is gated by the os_create_virtual_session_mode permission setting.

By default this is set to ask_each_time - the agent must stop and wait for your approval before launching anything.

You see the live feed of every session in the agent panel. Full transparency. If anything looks wrong, you can stop it immediately.

Workspace Boundaries

The agent operates within your project directory by default. It cannot read, write, or delete files outside your workspace unless you explicitly enable outside access in Settings > AI.

Files matching .agentignore or .gitignore patterns are excluded by default, protecting build artifacts, secrets, and other sensitive files from being read.

Your API Keys Are Local

API keys are stored in %APPDATA%\Roaming\zerosphere-nightly\.env on your machine with restricted permissions to prevent unauthorized access. They are never transmitted to ZeroSphere servers. Every request goes directly from your machine to your chosen model provider.

ZeroSphere is not in the loop between you and your AI provider.

Summary

What the agent can seeWhat the agent cannot see
Pixels rendered on screenDOM structure or hidden HTML
Your application's visible UIAccessibility tree or element IDs
Files within your workspaceFiles outside your project directory
Commands you approveAnything blocked by your permission settings