Skip to content
Cloud Security Risk Assessment

The 40-question assessment: how the cloud access score is calculated

The method: 40 questions in six domains, answers worth 1, 0.5 or 0, weights from 1 to 3, a domain score from 0 to 100 and five fixes chosen by the largest weighted gap, with ties broken by lower effort. This page shows the arithmetic, the domain weights and the sources for the controls.

Start the assessment ยท Download the questions (CSV)

Each of the 40 questions has three answers: yes in every account, in some accounts, and no or not known. They are worth 1, 0.5 and 0. A question carries a weight of 1, 2 or 3 that records how much a wrong answer moves risk. The 40 weights add up to 84 and 12 questions weigh 3.

A domain score is the sum of value times weight over the domain's questions, divided by the sum of weights, times 100. The six domains hold 8, 7, 7, 6, 6 and 6 questions. Domain weights of 25, 20, 20, 15, 10 and 10, which add up to 100, apply only to an optional overall figure. Unknown counts as no, so a missing control and an unprovable one score the same.

The five fixes are the questions with the largest weight times one minus the answer value. Ties go to the lower fix effort, rated low, medium or high, and then to question order. Every question names the control it comes from, and every question page gives a test for AWS, Azure and Google Cloud, so a score can be checked rather than trusted.

What each answer is worth

QuestionDomainWeightAnswers and valuesFix effort
1. Does the root user (or Global Administrator, or Organization Admin) of every account have phishing-resistant MFA and no access keys?Identities and entitlements3Yes, in every account = 1; In some accounts = 0.5; No or not known = 0low
2. Do all people reach the cloud through central single sign-on with MFA, with no local IAM users or standalone cloud passwords?Identities and entitlements3Yes, all people use SSO with MFA = 1; Mostly, with some local users = 0.5; No or not known = 0medium
3. Are long-lived access keys and service account keys absent, or rotated at least every 90 days and unused ones deleted?Identities and entitlements3Yes, none or all under 90 days = 1; Some are older or unused = 0.5; No or not known = 0medium
4. Do workloads get cloud access through attached roles, managed identities or workload identity federation instead of stored secrets?Identities and entitlements2Yes, for all workloads = 1; For some workloads = 0.5; No or not known = 0high
5. Is the number of people with administrator-level roles small, named and separate from their daily accounts?Identities and entitlements2Yes, small, named and separate = 1; Small but not separate = 0.5; No or not known = 0low
6. Are privileged roles granted only for a limited time, on request, instead of held permanently?Identities and entitlements3Yes, for all privileged roles = 1; For some privileged roles = 0.5; No or not known = 0high
7. Are permissions reviewed at least every 90 days and unused ones removed?Identities and entitlements2Yes, reviewed and removed = 1; Reviewed, but removal is not enforced = 0.5; No or not known = 0medium
8. Are emergency (break-glass) accounts in place, protected, monitored and tested at least twice a year?Identities and entitlements1Yes, protected, alerted and tested = 1; Exist but untested or unmonitored = 0.5; No or not known = 0low
9. Is every SSH, RDP and database port closed to the whole internet (0.0.0.0/0 and ::/0)?Network reachability3Yes, none are open = 1; Open on a few exceptions = 0.5; No or not known = 0low
10. Do your data stores and platform services (storage, databases, secrets, AI) accept traffic only through private endpoints?Network reachability2Yes, for all production services = 1; For some services = 0.5; No or not known = 0high
11. Do new compute instances and databases get no public IP address by default?Network reachability2Yes, enforced by policy = 1; By convention, not enforced = 0.5; No or not known = 0medium
12. Do administrators reach servers through a brokered, logged service (Session Manager, Bastion, Identity-Aware Proxy) with no inbound admin ports?Network reachability2Yes, for all servers = 1; For some servers = 0.5; No or not known = 0medium
13. Is the instance metadata service hardened (IMDSv2 required on AWS, managed identity endpoint protected on Azure, metadata concealment on GKE)?Network reachability3Yes, enforced everywhere = 1; On new resources only = 0.5; No or not known = 0low
14. Is outbound traffic from production workloads filtered to known destinations?Network reachability1Yes, deny by default with an allow list = 1; Partly, for some networks = 0.5; No or not known = 0high
15. Do you analyze which resources are reachable from the internet and from other networks (peering, transit, cross-account) at least monthly?Network reachability1Yes, on a schedule with owners = 1; Ad hoc = 0.5; No or not known = 0medium
16. Is public access to object storage blocked at the organization level, not just per bucket?Data perimeter3Yes, organization-wide = 1; On most accounts or buckets = 0.5; No or not known = 0low
17. Do your resource policies and sharing rules limit access to identities in your own organization?Data perimeter3Yes, enforced by policy = 1; In some policies only = 0.5; No or not known = 0medium
18. Can your most sensitive data be read only from trusted networks or endpoints (VPC endpoints, service perimeters, private links)?Data perimeter2Yes, for all sensitive data = 1; For some data stores = 0.5; No or not known = 0high
19. Do you review findings about resources shared outside your organization at least weekly?Data perimeter2Yes, weekly or continuous with owners = 1; Occasionally = 0.5; No or not known = 0low
20. Are sensitive data encrypted with customer-managed keys whose use is limited to named identities and services?Data perimeter2Yes, for sensitive data = 1; For some data = 0.5; No or not known = 0medium
21. Do you know which storage locations hold regulated or sensitive data, and is that list up to date?Data perimeter1Yes, scanned and tagged = 1; From memory or partly = 0.5; No or not known = 0medium
22. Is every third-party and cross-account role protected by an external ID or condition, with an owner and an end date?Data perimeter2Yes, all have conditions, owners and dates = 1; Some do = 0.5; No or not known = 0medium
23. Does each AI agent, pipeline and application have its own cloud identity, not a shared one or a person's credentials?AI workloads3Yes, one identity per agent or application = 1; Some share an identity = 0.5; No or not known = 0medium
24. Is the list of actions each AI agent can perform limited to what its task needs, and written down?AI workloads3Yes, scoped and reviewed = 1; Partly scoped = 0.5; No or not known = 0high
25. Can only approved models and AI services be used in your accounts, enforced by policy?AI workloads2Yes, enforced by policy = 1; By guidance only = 0.5; No or not known = 0low
26. Is the data an AI system can read (retrieval stores, training data, tool outputs) limited to what its users may see?AI workloads2Yes, access follows the user = 1; Scoped to the data store, not the user = 0.5; No or not known = 0high
27. Do destructive or high-impact agent actions (delete, grant access, pay, send externally) need a human approval?AI workloads2Yes, enforced in the tool = 1; Documented, not enforced = 0.5; No or not known = 0medium
28. Are model calls, tool calls and the identity behind each logged and kept?AI workloads2Yes, logged and retained = 1; Partly logged = 0.5; No or not known = 0low
29. Do organization-level policies block the riskiest actions before they happen (SCPs, Azure Policy deny effects, organization policy constraints)?Guardrails and drift3Yes, deny rules cover our top risks = 1; A few deny rules = 0.5; No or not known = 0medium
30. Do infrastructure changes go through code review and pipelines, with console changes in production flagged?Guardrails and drift2Yes, console writes are flagged = 1; Mostly code, console not monitored = 0.5; No or not known = 0high
31. Does a continuous check compare live configuration against your rules and report differences within a day?Guardrails and drift2Yes, continuous with owners = 1; Periodic scans = 0.5; No or not known = 0medium
32. Do you test a new guardrail in audit or report-only mode and see what it would block before you enforce it?Guardrails and drift1Yes, always = 1; Sometimes = 0.5; No or not known = 0medium
33. Does every guardrail exception have an owner, a reason and an expiry date?Guardrails and drift1Yes, tracked with expiry = 1; Tracked, no expiry = 0.5; No or not known = 0low
34. Can administrators inside a workload account not turn off logging, guardrails or security services?Guardrails and drift2Yes, protected from account admins = 1; Protected for some controls = 0.5; No or not known = 0medium
35. Are control-plane audit logs on in every account, subscription and project, in every region, including data access logs for sensitive stores?Evidence3Yes, everywhere = 1; In most places = 0.5; No or not known = 0low
36. Are audit logs stored in a separate, restricted account where workload administrators cannot change or delete them?Evidence2Yes, separate and write-once = 1; Separate, but not write-once = 0.5; No or not known = 0medium
37. Do you keep audit logs for at least 365 days, with the most recent 90 days searchable?Evidence2Yes, 365 days or more = 1; Under 365 days = 0.5; No or not known = 0low
38. Do alerts fire, and reach a person, for root or global administrator sign-ins, policy changes, new keys and disabled logging?Evidence2Yes, all four route to on-call = 1; Some are alerted = 0.5; No or not known = 0medium
39. Within one working day, can you produce a list of who can access a given sensitive data store and how they got that access?Evidence1Yes, within a day = 1; Within a few weeks = 0.5; No or not known = 0medium
40. Have you practiced cutting off all access of one compromised identity within one hour, and written down the steps?Evidence1Yes, in the past 12 months = 1; Planned, not practiced = 0.5; No or not known = 0high

Twenty known cases

The scoring is checked against twenty cases kept in the site's tests folder, written by hand from the model above; the build fails if one gives a different answer.

Cite this line: 'Cloud Security Risk Assessment, the 40-question assessment, edition 16 Nov 2026, https://cloudsecurityriskassessment.com/methodology/'

Sources

  1. CIS AWS Foundations Benchmark v5.0 1.4 and 1.5, read 16 Nov 2026
  2. Microsoft cloud security benchmark IM-6, read 16 Nov 2026
  3. Google Cloud security best practices Service account keys, read 16 Nov 2026
  4. Microsoft cloud security benchmark IM-3, read 16 Nov 2026
  5. Microsoft cloud security benchmark PA-1, read 16 Nov 2026
  6. Microsoft cloud security benchmark PA-2, read 16 Nov 2026
  7. Microsoft cloud security benchmark PA-4, read 16 Nov 2026
  8. Microsoft cloud security benchmark PA-5, read 16 Nov 2026

Last reviewed 16 Nov 2026.