Platform

Runtime guardrails that Glacis runs for you.

Your team gets AI into production without building a control layer first. Each governed action is checked against agreed controls while it is live, and the ones that cross them are narrowed, held or refused. Every covered decision leaves a signed record an auditor, a customer’s security team or an underwriter can check for themselves.

How the risk loop works

Decisions made while the action is live

Actions that stay inside the agreed controls go straight through. Glacis steps in only when one is crossed, and it makes that call while the action is live rather than in a queue someone reviews later.

  1. 01

    AI attempts an action

  2. 02

    Context and bounds

  3. 03 · Decision
    Allow Constrain Hold Deny
  4. 04

    Outcome reported

  5. 05

    Signed record

Allow: the action goes through unchanged. Constrain: it goes through with the risky part narrowed or redacted. Hold: it waits for a person. Deny: it’s refused, and the refusal is recorded.

Build versus buy

Nine things you don’t have to build

This is what sits between an agent that works and an agent anyone will let near production. Glacis runs this layer as a managed service, through every model swap and every new rule. None of it lands on your roadmap.

  1. 01Agent and workload identity
  2. 02Scoped credentials
  3. 03Network and tool boundaries
  4. 04Policy evaluationLive
  5. 05Classifier or model-based review
  6. 06Isolation
  7. 07Escalation
  8. 08Versioned evidenceLive
  9. 09Incident reconstructionLive

Policy evaluation, versioned evidence and incident reconstruction are in the product today. The other six stay the systems you already run, and Glacis works inside them. It doesn’t replace your identity provider, your cloud or your security program.

After guardrails

The questions the record already answers

The record is already there when the questions come: the sign-off before an agent touches a real system, the security review a customer runs before they buy, the question months later about one specific action. It builds while the AI works, so nothing gets assembled under a deadline. Guardrails come first, because outside review and risk transfer both need a record to work from.

  1. 01

    Guardrails, then evidence

    Glacis sits between the application and the models it calls. A record of each covered decision builds as it runs, one record per governed action, so the covered set can be counted.

    Glacis

  2. 02

    Outside review

    An independent evaluator examines the controls and the record. Where a recognized certification applies, the record is one input to the assessor’s process.

    Independent evaluator · accredited assessor

  3. 03

    Risk transfer

    You can put a defined risk schedule to insurers, usually through a broker. Because each covered decision is one record, the exposure can be counted, and the count can be checked against the records without access to your systems. Underwriting, wording, pricing and claims are theirs to decide.

    Broker · underwriter · licensed insurer

Glacis is not a certifier or an insurer, and doesn’t guarantee certification, insurability, pricing or claims outcomes. Steps two and three are what others can do with the record, not services Glacis performs.

Let’s look at where this would sit in your stack — talk to us.

Operator documentation: docs.glacis.io.