SANDBOXES FOR SUPERINTELLIGENCEIN DEVELOPMENT

Know what your agent can reach.

We’re building practical containment tests for coding agents. Join the waitlist for early access and project updates.

Drag to rotate · Shift + drag to roll

Drag to rotate the view around X and Y. Hold Shift and drag to rotate around Z. Use the arrow keys in the scene to rotate five degrees, Shift with Left or Right to roll, and Home to reset. The three sliders below also control the view. The ball stays inside the cube.

CAPABILITY, CONTAINEDILLUSTRATION / 001
CAPABILITY NEEDS CONTAINMENT.EXPLORE THE IDEA

/ THE APPROACH

A boundary should be
something you can test.

As agents gain access to more of our systems, their limits need to be understood. That’s where we’re starting.

01

Map access.

Understand the files, credentials, tools, and networks within your agent’s reach.

PERMISSIONS / VISIBILITY
02

Test boundaries.

Check whether the limits you set hold up in controlled, repeatable scenarios.

CONTAINMENT / VERIFICATION
03

Keep evidence.

Turn observations into clear findings and checks you can run as your setup changes.

FINDINGS / REPEATABILITY

FROM THE BEGINNING

Early days.
Clear direction.

Sandboxes is an independent project exploring agent containment. We’re looking for people building with coding agents to help shape what comes next.

Get the first update