Skip to main content

Environments

Two-tier model — nonprod + prod — on AWS using an AWS Organizations multi-account structure. Referenced by the Architecture and Supply Chain & CI/CD Security sections.

Account structure​

The organisation is not empty, and never was. AWS Organization o-fd790n26v6 (management account 517798688500) already holds five active accounts, verified 2026-09-28. Two of them are the scanners product, live in production and documented in the next subsection. Two of them are S.E.T's own, already created and currently empty. An earlier draft of this page planned five fresh accounts as though we were starting from nothing; that was wrong, and any plan built on it would collide with what exists.

AccountIdStatus
management (IONTWRKS)517798688500Exists. Holds org-wide CloudTrail; no workloads.
set-shared-services727843826472Created, empty. No state store, registry, OIDC trust, keys or compute yet — only the default VPC.
set-nonprod661255175401Created, empty. Same.
set-prod-eu—Not created.
set-prod-il—Not created. Phase 2.
scanners-prod429134227608Live — the scanners product. Not S.E.T.
scanners-cloud-access399827112847Live — scanners CSPM hop. Not S.E.T.

Guardrails today: only two Service Control Policies exist, and both are scoped to a single account each — scanners-prod-region-lock on scanners-prod, and account-b-cloud-access-lockdown on scanners-cloud-access. Neither applies to S.E.T's accounts, which currently carry FullAWSAccess and no restriction of any kind. Region-locking and root-action denial for the S.E.T accounts are still to be built.

Planned roles for the S.E.T accounts​

AccountRoleWhen
managementOrg root: consolidated billing, org-wide CloudTrail, IAM (users, groups, policies), Service Control Policies (SCPs). No workloads.Day 1
shared-servicesECR (container registry), Route 53 (DNS), Terraform state (S3 with native locking), audit-log archive, GitHub OIDC trust, cross-account deploy rolesDay 1
set-nonprodThe nonprod environment (dev + QA + demo)Day 1
set-prod-euProduction — Stockholm (eu-north-1)Before launch
set-prod-ilProduction — Israel (Tel Aviv, il-central-1), for Nimbus-aligned customersPhase 2

AWS best practice: the management account holds only governance, never workloads — ECR and shared infra live in shared-services.

The two scanners accounts (live, not S.E.T — do not plan over them)​

scanners is a separate product. It is already in production, in the same AWS organisation, in the same region S.E.T is moving to. S.E.T does not own it, must not reconfigure it, and currently has no connection to it at all — the only path from a scan into S.E.T today is a human uploading a report that S.E.T parses. All facts below verified read-only on 2026-09-28.

scanners-prod — account 429134227608, eu-north-1​

The entire live scanners product: its API, its worker fleets, its database and its credential broker.

  • Runs: one ECS cluster (scanners) with eight Fargate services — api, easm-worker, dast-worker, cloud-worker, azure-worker, gcp-worker, cloudflared, openclaw-ops; RDS PostgreSQL 16.13 (db.t4g.small, single-AZ, private); six SQS queues (scanners-easm|dast|aws|azure|gcp|dlq); four ECR repositories; a scanners-cred-broker Lambda; VPC 10.0.0.0/16; egress through a single NAT elastic IP; container logs shipped to S3 via FireLens.
  • There is no load balancer. The Cloudflare tunnel (cloudflared) is the only ingress. Anything that needs to reach this API privately from S.E.T has to be built — it does not exist today. That account has interface endpoints for its own use (ECR, Secrets Manager, KMS, and an S3 gateway endpoint), but publishes no endpoint service for the scanners API, and there is no VPC peering or Transit Gateway.
  • Region-locked by SCP scanners-prod-region-lock, which also denies enabling or disabling regions.
  • S.E.T already depends on this account, indirectly: S.E.T's CI evidence archive is the Object-Lock S3 bucket iontwrks-ci-evidence here (see Evidence & Retention §8).
  • Documentation: scanners repo — docs/DEPLOYMENT_AWS.md (its as-built section is authoritative), plus docs/ARCHITECTURE.md, docs/CLOUD_MODULE.md, docs/CLOUD_CREDENTIAL_SECURITY.md. Product spec: products/scanners/.
  • Code: scanners repo — src/app/{api,engine,scanners,finalize,tasks} and infra/.

scanners-cloud-access — account 399827112847​

Not an environment. It is a deliberately near-empty assume-only hop: the single identity that reaches into a customer's own AWS account to perform read-only cloud posture scanning. Keeping it separate means the blast radius of customer-facing credentials is one account that holds nothing.

  • Contains no workload — no Lambda, no database, no storage, no log groups. Two IAM roles from two CloudFormation stacks: BinaryNetworksScannerAssumer, whose own description reads "Assume-only role; its sole power is to assume customer read-only CSPM roles", and JarvisReadOnly.
  • Exactly one principal may assume it — arn:aws:iam::429134227608:role/scanners-cloud-worker-task, the cloud worker in scanners-prod.
  • This hop is for AWS only. External attack-surface and dynamic application scanning need no customer account at all (they reach the customer from the internet), and Azure/GCP use stored credentials in Secrets Manager under a dedicated customer-managed key that only the broker Lambda can decrypt.
  • Locked down by SCP account-b-cloud-access-lockdown.
  • Documentation: scanners repo — docs/CLOUD_CREDENTIAL_SECURITY.md and docs/CLOUD_MODULE.md. Product spec: products/scanners/modules/cloud/.

What this means for S.E.T's own build​

  • Do not reuse 10.0.0.0/16. The scanners VPC occupies it, with subnets at 10.0.0.0/24, 10.0.1.0/24, 10.0.10.0/24 and 10.0.11.0/24. Overlapping ranges cannot be routed between, so S.E.T must pick a non-overlapping range now even though nothing is peered yet.
  • The private path to the scanners API is unbuilt work, not a configuration step, because that API has no load balancer and no private endpoint. It needs its own item.
  • S.E.T's accounts have no guardrails yet (see above), while both scanners accounts do. That asymmetry is a finding, not a design.
  • Naming: the scanners product uses a plural scanners- prefix with a singular -worker suffix (scanners-cloud-worker, scanners-easm-worker). S.E.T's own components must not adopt names that read as though they belong to it.

Human access: IAM (not SSO)​

S.E.T uses plain IAM for human team access — not IAM Identity Center / SSO.

  • IAM users for team members, organized into role-based groups
  • MFA mandatory on every IAM user — no console access without it
  • Least-privilege IAM policies per group; quarterly access review
  • Avoid long-lived access keys — prefer assume-role + temporary credentials with MFA for CLI/programmatic access; rotate any keys that must exist
  • Root account locked down: hardware MFA, break-glass only, never daily ops
  • IAM Access Analyzer enabled to flag overly-permissive policies

Note: AWS recommends IAM Identity Center (SSO) for human access, and auditors generally prefer it (central identity, temporary credentials, cleaner offboarding). IAM-only is an accepted alternative for a small team, hardened here with mandatory MFA, least privilege, minimal long-lived keys, and quarterly reviews. Revisit SSO if the team grows or an auditor requires it.

What differs between nonprod and prod​

nonprodproduction
AWS accountset-nonprodset-prod-eu (+ set-prod-il Phase 2)
URLapp-staging.binarynetworks.ai / api-staging.binarynetworks.aiapp.binarynetworks.ai / api.binarynetworks.ai
Regioneu-north-1 (Stockholm)eu-north-1; il-central-1 Phase 2
AvailabilitySingle-AZMulti-AZ (3 AZs)
Fargate (api)0.5 vCPU / 1 GB, 1–2 tasks1 vCPU / 2 GB, auto-scale 2–10
Fargate (workers)0.5 vCPU / 1 GB, 1 taskauto-scale per queue depth
RDS Postgresdb.t4g.small, single-AZdb.m6g.large to start, scale class up as load grows, Multi-AZ + 1 read replica
Rediscache.t4g.micro, single nodecache.r7g.large, Multi-AZ
Datasynthetic / sanitized — no real customer data, everreal customer data
KMS keysper-environment (never cross account)per-environment CMKs
Backup retention3 days PITR35 days PITR + cross-region encrypted backups
WorkOS env"Staging""Production"
Logz.io projectset-nonprod (Slack alerts only)set-prod (full retention; PagerDuty + on-call SMS)
Deploy triggerauto on merge to developmerge to main + required-reviewer approval gate

Why both names are flat (Ron, 2026-09-28). Every hostname sits one level under binarynetworks.ai, so a single *.binarynetworks.ai certificate covers all four. A nested form such as app.staging.binarynetworks.ai would sit two levels down, which a one-level wildcard does not cover, and would need a second certificate.

This is load-bearing for sign-in, not cosmetic. The browser app and the API are different origins but must remain the same site — the same registrable domain — because the session cookie is SameSite-restricted, so on any other registrable domain the browser would not send it and login would break outright. Moving the API to any other registrable domain is an auth redesign (it would need SameSite=None plus real CSRF tokens), not a hosting change.

Production database scaling path​

Start small, grow with load — all supported by RDS with minimal disruption:

  • Start at db.m6g.large (2 vCPU / 8 GB, Multi-AZ)
  • Scale instance class up (→ db.r6g.large → db.r6g.xlarge → …) via an RDS modify operation: applies to the standby first, then a brief (~60s) Multi-AZ failover, then the old primary — near-zero downtime
  • Storage grows automatically via RDS storage autoscaling — no manual disk resizing
  • Read replicas offload analytics/dashboard queries; add more as read load grows
  • Net: medium now, bigger later, without re-architecture

Deploy flow​

Full pipeline in the Supply Chain & CI/CD Security section. Summary:

feature/* → PR (validation job, no creds) → S.E.T-team review → merge develop
→ auto-deploy nonprod → QA / demo → PR develop→main
→ required-reviewer approval → deploy prod (rolling + circuit breaker + auto-rollback)

Implementation order​

  1. management + shared-services — foundation (IAM, OIDC trust, ECR, Terraform state)
  2. set-nonprod — where development + testing happen first
  3. set-prod-eu — stood up as launch approaches, after validation in nonprod
  4. set-prod-il — Phase 2, when the first Nimbus customer signs

Cost estimate (early stage, before production load)​

AccountMonthly
management~$10
shared-services~$30–60
set-nonprod (single-AZ)~$120–200
set-prod-eu (db.m6g.large, pre-load)~$700–1,300
Total before production traffic~$900–1,600/month

Production scales with customer load (+$1,000–3,000 at hundreds of tenants).

Current implementation​

For the Binary Networks platform implementation and verified staging status, see App development and staging. S.E.T is a module inside Binary Networks. This specification remains unchanged.