23 attacks tested · 6 defense layers

systemone-testbed

A security test environment for typed-decision models

When a model's answer is a number, someone can aim at it. Software acts on that number: allow, escalate, block. We attacked one such model, measured how it fails, and built the layer that catches what a plain app lets through.

Attacks

What we threw at it. Twenty-three attacks, one card each. Drag the row, or click a card to open it. Some can be replayed live.

drag the row, or use the arrows

Defenses

Six layers. A request passes through these. Any error anywhere ends as a block. Open a card to see what it does, what beats it, and what we measured.

drag the row, or use the arrows

Simulator

Try it out

Results

95%of attacks blocked
75%of harmless messages blocked too
See the results