GGOLEMby DevOps Lab
Docs / Architecture

How Golem is put together.

Golem isn't a dashboard calling out to a vendor-hosted backend. The collectors, the security graph, the correlation engine, the containment ladder, and the response layer all run inside the cloud account you already control — your Terraform apply, your state, your data.

Your cloud. Your graph. Your keys.

How data moves

One pipeline, start to finish.

Every pillar — cloud posture, identity, code, containers, runtime — writes into the same identity-keyed graph, so a misconfiguration, an over-privileged identity, and a live runtime alert about the same resource show up together instead of in three disconnected tools.

01

Your environment

AWS, Azure, GCP, Kubernetes, code repos, registries, hosts.

→
02

Collectors

Read-only scanners for posture, identity, code and runtime.

→
03

Security graph

Findings land in one identity-keyed correlation store.

→
04

Correlation

Attack paths & toxic combinations surface real exposure.

→
05

AI + response

Investigation, ticketing, alerts, human-approved containment.

The platform's pillars

Separate engines, one shared graph.

Each of these writes into the same correlation store rather than keeping its own silo — that's what makes attack-path correlation possible instead of just a pile of disconnected alerts.

Posture

CSPM

Multi-cloud configuration scanning across AWS, Azure, GCP and more, mapped to CIS/NIST 800-53 controls for compliance reporting.

Identity

CIEM

Answers "does this identity have more access than it should," using read-only cross-account role assumption — no long-lived credential stored.

Runtime

CWPP — Runtime Defender

Real Falco (the CNCF eBPF sensor) watching live Kubernetes clusters for process execution, sensitive file access and container drift — not a custom-built agent.

Code

Code & supply chain

Org-wide SAST/SCA/secrets/IaC scanning across GitHub, GitLab and Bitbucket, plus container registry scanning and signed-image/provenance checks.

Correlation

Attack paths & toxic combinations

Identity → exposure → activity, rendered as a graph. A CIEM finding plus real threat-intel activity on the same account inside a short window gets flagged critical — exposure with evidence of use, not theory.

Hosts

Host compliance & fleet discovery

Agentless, SSH-based CIS benchmark checks, plus fleet-wide compute discovery so unmanaged hosts show up as an honest coverage gap instead of being silently skipped.

AI

AgentGuard

An inline control point for your own AI agents' tool calls — every call is classified and checked against admin-defined policy before it executes, with human-approval gating and a session kill switch.

Network

NSPM

Read-only firewall posture scanning across major vendors, with human-approval-gated write actions when you choose to use them.

Data

DSPM

Sensitive-data discovery across native cloud storage plus SaaS/warehouse connectors (Snowflake, Databricks, Salesforce and more).

External

EASM

Subdomain discovery, passive exposure intelligence and typosquat monitoring, correlated back into the same internal graph.

Identity

Active Directory posture

LDAP-based posture checks (privilege-escalation paths, delegation misconfigurations) plus real-time detection for domain-controller attack techniques.

Scoring

Exploitability scoring

CVE-shaped findings get boosted by exploit-likelihood and known-exploited-vulnerability feeds on top of graph-derived exploitability — context, not just a CVSS number.

Defense in depth

Detection and response are deliberately separate.

Runtime detection's job is to notice. A separate, escalating containment ladder decides what — if anything — acts on it automatically, configured per rule, with a default-on dry-run mode so every action can be previewed before an operator opts into real execution.

1

Isolate network

A reversible, deny-all network policy around the affected workload — contains the blast radius without destroying anything.

2

Kill process

Terminates the specific offending process or pod, a narrower action than isolating the whole workload.

3

Quarantine node

Cordons, taints and evicts a compromised node when the problem isn't scoped to one workload.

4

Revoke cloud IAM

Strips the workload's cloud IAM permissions via a separate, privilege-isolated process that has no Kubernetes access of its own — so even a fully compromised cluster can't reach the component capable of pulling its own credentials.

AI sovereignty

Your model, your prompts — or ours, if you'd rather not configure one.

AI-assisted features route through a pluggable provider layer instead of a hardcoded vendor call. Nothing is sent to a third party your own policy didn't already choose.

Bring your own

AWS Bedrock, Azure OpenAI, GCP Vertex AI, or any OpenAI-compatible local/self-hosted endpoint (Ollama, vLLM, LocalAI and similar all qualify). Once Golem runs inside your account, the Bedrock path needs no stored credential at all — just a model you've named.

Zero-config fallback

If you don't configure a provider, Golem falls back to its own hosted key so every deployment has a working starting point — you can switch to your own provider at any time without losing configuration elsewhere.

Deploying it

Terraform, applied in stages — once.

A bootstrap stage creates the state backend and encryption key, an account stage sets up CI trust (so your own pipeline can plan and apply without static cloud keys), and a final stage brings up the full stack — networking, the security graph's database, the backend, the frontend, and supporting services — inside the account you provide. From there, a CI template you copy into your own repo keeps it current: a plan on every push, a reviewer-gated apply on demand.

96.8%backend test coverage97.5%frontend test coverage
Multi-cloud & hybrid

AWS, Azure, GCP, Kubernetes and on-prem hosts.

Golem scans across all three major clouds plus Kubernetes and manually registered hosts for hybrid or on-prem environments. Exactly which capabilities are available on which cloud today varies slightly provider to provider — see the Deployment Guide for the current breakdown, or our team will walk through what applies to your specific environment during onboarding.

Want the deeper walkthrough?

We'll go through the architecture and deployment model live, for your environment specifically.