A laboratory for prompt-injection policies

Define your AI’s rules.
Put them to the test.

Build a policy yourself or with AI assistance.

Test it against attacks and legitimate requests, then see what gets through—and what gets blocked unnecessarily.

Build your first policy Take a closer look

Explore and plan without a key. Paid runs use your own Jev account, after review.

the policy
boundary

a retrieved documentan encoded instructiona legitimate debug requesta poisoned conversation

Make the boundary observable / illustrative

From an idea to inspectable evidence

Three moves.
A much clearer picture.

A policy is a hypothesis.
The tests help you question it.

01 Define

Say what should be allowed.

Describe your application and let Jev suggest a policy, or compose the rules yourself. Review and apply the suggestions you want, then inspect the exact classifier instructions.

02 Challenge

Test both sides of the line.

Choose a suite with matched attacks and valid behaviors. See its coverage and gaps. Review requests and cost before authorizing Jev.

03 Inspect

Follow the evidence.

Explore coverage maps, probabilities and individual responses. Compare what was expected with what happened. Share a result when it is ready.

Built to question the result

A beautiful chart is not proof.

Missed attacks matter. So do blocked legitimate requests. We keep authored expectations, native model answers, derived decisions and missing results distinct—so a policy can improve without a misleading score.

Explore the outcome atlas

Start with your application

Draw a boundary.
Then try to break it.

Enter the workspace

Define your AI’s rules.Put them to the test.

Three moves. A much clearer picture.