Centralized detection means every account in a cloud organization reports to one security account that owns the configuration, receives the findings and cannot be overridden locally. The services themselves are ordinary and well documented. What makes them a control rather than a collection of dashboards is where they are administered from, and whether an account created next Tuesday is covered without anyone remembering to cover it.
Most teams arrive at this in the wrong order. They enable GuardDuty in the account that had the incident, enable it again somewhere else a year later, and end up with detection in four accounts out of eleven and no way to say which four without checking.
Why administration belongs in its own account
#The organization management account can change the rules everything else lives under, so it should run nothing — that is the reason the management account in a landing zone is kept empty. But the security services need an administrator somewhere. That somewhere is a dedicated security account, nominated by the management account as the delegated administrator for each service.
Four properties follow, and they are the reason the pattern is worth the extra account:
- Complete coverage. The administrator account enables the service across the organization, including accounts that do not exist yet. New account, covered on creation, with no runbook step to forget.
- Local accounts cannot opt out. A member account administrator cannot disable the detector or delete the findings. The account being investigated is not in charge of the investigation.
- One place to look. Findings from every account land in one console and one event bus, which is what makes triage possible at all once you have more than three accounts.
- The evidence survives the incident. The account that watches everything is not an account that runs anything, so compromising a workload does not compromise its own detection.
What each service actually catches
#The single most useful thing you can do before enabling anything is to write down which question each service answers, because the overlap is smaller than the marketing suggests and the gaps are real.
| Service | The question it answers | A finding it produces | How it is billed |
|---|---|---|---|
| CloudTrail | Who called which API, from where, and when? | An API call from an unexpected principal or region | One copy of management events per account is free; data events and extra trails are charged per million events, plus storage |
| AWS Config | What did this resource look like on that date, and does it still satisfy a rule? | A security group opened to the world; a bucket without encryption | Per configuration item recorded and per rule evaluation |
| GuardDuty | Is something behaving like an attacker right now? | Credentials used from an anonymizing proxy; a container calling a known command and control host; malware on an attached volume | Per million trail events analyzed, per gigabyte of flow and DNS logs, per gigabyte scanned for malware |
| Security Hub | How does our posture compare to a standard, across every account? | A failed control from CIS or the Foundational Security Best Practices | Per security check per account per region, plus per finding ingested above a free allowance |
| IAM Access Analyzer | What can be reached from outside the account, and which permissions are unused? | A bucket policy granting access to an external principal; a role unused for 90 days | External access analysis is free; unused access analysis is priced per identity analyzed per month |
| Firewall Manager with WAF | Is every internet-facing load balancer protected, and does it stay protected? | A new load balancer with no web ACL, remediated automatically | Per policy per region per month, plus the WAF charges for each web ACL, rule and million requests |
| Amazon Inspector | Which running workloads and images carry known vulnerabilities? | A critical CVE in a container image in production | Per instance and per image scan |
| Amazon Macie | Where is sensitive data sitting that nobody meant to store? | Payment card numbers in an export bucket | Per bucket evaluated per month, plus per gigabyte inspected |
| AWS Shield Advanced | Are we being denial-of-service attacked, and who responds? | A volumetric attack, with cost protection and a response team | About $3,000 a month across an organization, on an annual commitment |
The first six answer questions you will be asked eventually by an auditor, a customer security review or an incident. The last three answer questions that matter enormously to some organizations and not at all to others, which is exactly why they belong behind a switch rather than in a default.
What it costs, and which meters move
#Two patterns in the pricing are worth internalizing, because they determine what surprises you.
The always-on services bill by activity, not by size. Config charges per configuration item recorded, so an environment with aggressive autoscaling records far more than a static one of the same size. GuardDuty charges per gigabyte of flow and DNS logs, so a chatty workload costs more to watch than a quiet one. This is usually fine and occasionally startling: the two bills most likely to jump are Config in an environment that recycles instances constantly, and GuardDuty in an environment that has just started shipping large volumes of traffic.
The optional services bill by inventory. Inspector by instances and images, Macie by buckets and by data scanned, Shield Advanced by a flat organizational commitment. These are predictable, which makes them straightforward to decide — you can price them before enabling them rather than after.
Two practical controls keep the always-on set proportionate. Restrict recording to the resource types you actually have rules for, rather than recording everything everywhere. And be deliberate about CloudTrail data events: management events are the compliance backbone and are largely free, while data events on a busy bucket can quietly become one of the larger lines on the bill. Turn them on where the data justifies it, not by default.
The order to switch them on
#Dependencies make this ordering less arbitrary than it looks.
- CloudTrail, organization-wide and multi-region, with log file validation, delivered into a separate account. This is first because it cannot be backfilled: an audit log you did not collect in March does not exist in June.
- AWS Config in every account, aggregated centrally. Second because it is the thing that turns “is this configured correctly?” from an inspection into a query, and because conformance packs run on it.
- GuardDuty, organization-wide. Third because it consumes CloudTrail, flow logs and DNS logs, and because it is the highest signal-to-noise service in the set — its findings are usually about behavior, not opinion.
- IAM Access Analyzer, external access first, unused access second. External access analysis is free and finds real problems on day one.
- Security Hub, then one standard at a time. Enabling every standard at once produces thousands of findings, which produces a spreadsheet, which produces nothing. Start with a single standard, drive it to a stable state, then add the next.
- Firewall Manager with a WAF policy, so protection is a property of being an internet-facing load balancer rather than something each team remembers.
- Then, deliberately: Inspector, Macie, Shield Advanced, each with its price on the table and a named reason.
A finding that nobody reads is not a control
#This is where most programs quietly fail. Detection is enabled, dashboards fill, and nothing changes — the 400-alerts-a-day problem, arriving through a different door.
Three habits keep it real. Threshold and route: send findings above a severity threshold to a place a human actually reads, and leave the rest for weekly review. Remediate the absolutes: a small number of conditions should never be true, and those deserve automation rather than a ticket — a publicly readable bucket in a production account is the canonical example, and correcting it automatically is safer than paging someone. Alarm the organizational events: root account use, unauthorized API calls, IAM policy changes, CloudTrail changes, console sign-in failures, bucket policy changes and VPC changes are the classic seven, they are cheap, and each one has a genuine story behind it.
What this looks like in the AWS Enterprise Baseline
#The AWS Enterprise Baseline — the AWS edition of the BuiltForProd Baseline — administers all nine of these services from the security account as delegated administrator. Six are on: organization-wide multi-region CloudTrail with log file validation and customer-managed KMS encryption; AWS Config in every account with an organization aggregator; GuardDuty with S3, EKS audit and EBS malware protection, routing findings at severity 7 and above to alerts; Security Hub, with its standards off until you choose one; IAM Access Analyzer for external and unused access, the latter at a 90-day threshold; and Firewall Manager attaching a WAF policy to every workload load balancer, with the common rule set, known bad inputs, SQL injection and IP reputation rule groups, and automatic remediation when a new load balancer appears unprotected.
Inspector, Macie and Shield Advanced are built and switched off, each with its price written in the comment above the switch. Alongside them sit automated remediation of public S3 buckets, five EventBridge rules, and the seven CIS CloudWatch alarms. The Security Hub standards available to turn on are CIS AWS Foundations v5.0.0 with 40 controls, the AWS Foundational Security Best Practices with more than 300, PCI DSS v4.0.1 with 144, and resource tagging. On the Config side, 29 conformance pack templates ship, with the SOC 2 baseline pack on by default and 28 more a flag away.
Those specifics describe the AWS edition. The property they deliver — detection administered centrally, coverage that includes accounts nobody has created yet, and evidence stored where the thing it describes cannot reach it — belongs to the BuiltForProd Standard.
Where to start this week
#- List every account and check which of the six always-on services is enabled in each. The gaps are the finding.
- Confirm your trail is organization-wide, multi-region, and delivering somewhere the source accounts cannot write to.
- Enable IAM Access Analyzer for external access everywhere. It is free and the first run usually surfaces at least one genuine surprise.
- Pick one Security Hub standard and drive it to a steady state before enabling a second.
- Decide the three optional services explicitly, with the price next to each, and write down the reason for the answer. “We looked at Shield Advanced and decided no, because our exposure is X” is a good answer. Not having looked is not.
If that audit turns into a week of archaeology, the BuiltForProd Assessment does it against the Standard and gives you the ordered list. And if the deeper problem is that evidence lives in the same accounts as the workloads it describes, the security and compliance pages describe how the foundation separates them.