Security & privacy
Honest about what a mandate protects – and what it doesn’t.
What a mandate protects – and what it doesn’t
A mandate does not make an AI agent smarter or more honest. It limits what the agent can do, whatever it has read or been told. That is why it helps against prompt injection, and why it cannot help where the guard itself is bypassed.
What is protected
- A manipulated agent can only do what the mandate allows.
- Critical actions need an explicit rule or a person.
- Every decision is logged. A change in the middle of the log is noticed; with signed checkpoints, changes anywhere are.
- Mandates are plain files you can read and check.
What is not
- A hacked home guard or hub.
- Mistakes in your own rules – an allow is an allow.
- Devices that can be controlled without the guard.
- What the agent’s provider does with your conversations.
Where the decision is made
Decided outside the agent
The guard that checks requests is not part of the agent. In the intended setup, mandates and logs are stored and evaluated on a device in your home, and only confirmation questions go to your phone. Nothing is sent to us. This website itself uses no cookies and no tracking. How the guard decides
Reporting a vulnerability
Found a vulnerability?
Please report it privately on GitHub. We acknowledge every report within 72 hours.
How to report(opens another website in a new tab)