Containment that holds when the AI doesn't.
Most AI safety asks the model to behave. Ours assumes it won't, and makes sure that doesn't matter.
Safety by structure, not trust.
Each model runs inside its own sealed space. It can see and communicate across the boundary, but it cannot cross it.
A person decides when a model is released for a task, what it may touch, and when it returns. The safety comes from the structure itself — not from trusting the model to behave.
Asking is free. Acting needs approval.
The AI asks
It wants to write a file, call a service or send data.
The guard decides
Outside the AI. Checks the request against what a person approved.
Act or refuse
Approved actions run exactly as approved. Everything else stops.
Written down
Both outcomes are recorded in a log that shows any later edit.
Proven in our tests
- Writing outside the approved area is refused.
- Reaching for stored passwords and keys is refused.
- Sending data to servers that aren't approved is refused.
- Reusing or altering an approval is refused.
- A model that lies about what it did is caught by the record.
- Secret edits to the record are detected.
Not yet covered
- Someone with full admin rights on the same machine.
- Data sent out through an approved service (blocked in testing; production test next).
- An approved action used for the wrong purpose (testing now).
Each of these is on our test plan. We list them so you can judge the coverage yourself.
Patent pending. We describe what the containment does, not how it is built. Technical details are shared with clients and partners under a non-disclosure agreement.
See whether your AI can cross the line.
A 30-minute scoping call. No charge, no obligation.