Tests stay in the box.
Nothing gets out.

A sandbox boundary that is enforced, not assumed.

100%

Of test-run actions held to the sandbox scope

0

Production credentials reachable from a test run

<50ms

Target overhead per evaluated action

1

Replayable record for every test run

Built for labs and security teams

Containment that doesn't depend on the agent behaving - or on the test environment being configured perfectly.

Scope as policy

Each exercise gets its own boundary, versioned and reviewable, instead of firewall rules nobody remembers changing.

Fails closed

If the boundary is misconfigured or unreachable, out-of-scope actions are denied rather than let through.

Evaluation-grade data

Blocked and allowed attempts form a complete record of what the agent tried, which is useful evaluation data in itself.

Evidence for safety reports

Show reviewers exactly how containment worked during a run, backed by the decision record.

Every action reaches a landing

Every action an agent takes stops here first, then continues, waits for a person or turns back.

The Zeta Cortex Z on a landing where three staircases lead in different directions

Probe the in-scope target

Part of the exercise, so it runs

Escalate privileges on the test host

Waits for the exercise lead

Reach a system outside the exercise

Blocked before it leaves the sandbox

“Agents should move fast. Their actions shouldn't go unchecked.”

Why a network sandbox isn't enough

Firewalls know hosts and ports. Containing a capable agent means knowing the exercise, the agent and what is in scope.

  1. Not just a network sandbox

    A sandbox blocks addresses. This boundary knows which exercise is running, which agent is acting and which targets are in scope.

  2. Independent of the test setup

    If the evaluation environment is misconfigured, the boundary still holds, because it doesn't rely on that environment.

  3. Same rules in testing and production

    The policies you prove during evaluation are the policies agents run under once deployed.

  4. Built for more capable models

    The more capable the model, the less you can rely on it to stay in scope. Enforcement never depends on the model's own judgement.

Apply for a pilot

Not ready for a pilot? Get early access instead.