GGOLEMby DevOps Lab
Docs / Deployment

Applies into your own cloud account.

Golem ships as Terraform. You — or your platform team — run the apply, into infrastructure you already own. Golem's own operators never reach into your account; there's no button in a dashboard that grants itself access.

How deployment works

The rules every deployment follows.

You apply it, in your own account

Terraform runs under your own credentials, through your own change-management process if you have one. There's no cross-account trust relationship that isn't initiated from your side first.

Read-only by default

The baseline scanning grant is read-only. No module in the deployment grants write, modify or delete access to your environment by default.

No long-lived keys where avoidable

Where the cloud provider supports it, Golem uses short-lived, assumable-role style trust instead of static credentials — nothing to rotate, nothing that leaks from a stolen file.

Capabilities are opt-in

Deeper capabilities — identity analysis, fleet-wide host discovery, additional scanning — are separate, explicit opt-ins layered on top of the baseline, not a single all-or-nothing grant.

Trust model, per cloud

AWS, Azure and GCP each earn trust differently.

Every cloud follows the same shared rule — you apply it, it's read-only, Golem's operators never touch your account directly — but the actual trust mechanics differ because the clouds themselves differ.

AWS

Role assumption, no static key

One IAM role trusts Golem's backend to assume it — no access key, nothing to rotate. The broadest capability depth lives here today, including identity/entitlement analysis and fleet-wide host discovery, both opt-in on top of the baseline read grant.

GCP

Workload Identity Federation, no key file ever created

Your project learns to trust Golem's existing cloud identity directly and exchanges it for a short-lived token on every scan — deliberately built this way since many GCP organizations now block service-account key creation outright.

Azure

The one cloud where a secret is unavoidable

Azure's own workload-identity federation only accepts OIDC tokens from a registered issuer, and AWS's request-signing isn't an OIDC token — so there's no direct bridge the way GCP has. A rotated client secret is the honest trade-off here, not an oversight.

Scope: all three clouds default to scanning one account/project/subscription. AWS and GCP both support granting the same read access at an organization or folder level in a single apply; Azure can do the same at a management-group level, with one caveat — a required security-provider registration step has to happen per subscription either way, since Azure has no broader-scope equivalent for that specific step.

Kubernetes

EKS, AKS, GKE, or anything else — one path.

Kubernetes scanning authenticates off a standard kubeconfig, the same code path regardless of which cloud (or no cloud) runs the cluster. Telling Golem which managed provider it's talking to only unlocks small, additive extras — like enriching findings with the cloud's own logging data — never a prerequisite for scanning to work at all. Live-verified against real managed clusters on all three major clouds: real API discovery across the resource types each provider actually extends Kubernetes with.

Hosts & runtime

Agentless hosts, real eBPF for runtime.

Host compliance, any cloud

Registering a host for CIS-benchmark checks is pure SSH — hostname, port, and a credential. There's no cloud-provider concept in this path at all, so a VM on any cloud (or on-prem) works identically. Golem's backend needs outbound SSH reachability to the host, which is a network consideration on your side, not a per-cloud setup step.

Runtime Defender, cloud-agnostic

Installing runtime detection is a one-line script that deploys a standard, open-source eBPF sensor via Helm — the same chart on any standard Kubernetes cluster. The optional, separate containment-responder step (only needed for automated response, not detection) has exactly one capability that's AWS-specific today: resolving which cloud role to revoke from a workload's own pod identity. That gap is explicit, not silently papered over.

Deploying into a specific environment?

Let's go through your account structure and plan the apply together.