This sheet is past its review date. Deadlines or advice may have changed since 5 Nov 2026.
Cloud configuration: which settings let attackers in, and would you notice if one changed?
This month, run CISA's ScubaGear against Microsoft 365 and a free posture check against each cloud, and fix public storage, long-lived access keys, open app consent and open external access first. Then lock those settings with organization-level guardrails and re-run the same checks every week, so a changed setting shows up as a diff rather than as an incident.
If you only do five things
Block public storage at the top.
Enforce S3 Block Public Access for the whole AWS organization16, set
AllowBlobPublicAccessto false on every Azure storage account13 and enforcestorage.publicAccessPreventionin Google Cloud22. Data stores left online without access controls remain among the most common errors in the DBIR6.Replace long-lived keys with short-lived credentials.
AWS traces a large portion of security events to compromised long-term access keys17, and in late 2025 attackers broke in with AWS keys suspected to have leaked in 20225. Move people to federated sign-in and workloads to roles or managed identities, then delete the keys.
Turn the baseline into guardrails, then diff it weekly.
Enforce what must never change with AWS Organizations policies, Azure Policy and Google Cloud organization policies, then re-run the assessment weekly and compare runs1.
Adopt a baseline without breaking things
Pick one baseline per platform.
For Microsoft 365, SCuBA, or Baseline Security Mode, which every plan has8; for Azure, the Microsoft cloud security benchmark12; for AWS, Foundational Security Best Practices15; for Google Cloud, its security baseline constraints21.
Assess, and write down each exception.
Export every failing check. Record why each one stays, who owns it and when it ends; ScubaGear's YAML config holds these as exclusions, annotations or omissions1.
Check the impact, then fix.
Start with the policies CISA marks as BOD 25-01 requirements4, and turn on each Baseline Security Mode setting whose impact report shows zero impact8. For MFA settings, use the passkeys and MFA sheet.
Enforce at the top of each cloud.
In AWS, SCPs cap what roles can do, RCPs what resources allow, and declarative policies set defaults such as IMDSv21918. In Azure, assign deny policies at a management group. In Google Cloud, add the baseline constraints if your organization predates 3 May 202421. Test each on one account first.
Diff every week.
Compare each run with the last, for example with ScubaGear's
Invoke-SCuBADiff1. Scan infrastructure-as-code pull requests with Checkov or a similar checker23, and treat a scheduledterraform planthat isn't empty as drift.
Settings attackers go after
- Public storage. Google's public access prevention revokes existing public access once you enforce it22. Move public sites and downloads to a separate account or project first.
- Access keys and shared keys. Attackers reach Azure Blob Storage with exposed credentials, shared keys and SAS tokens7; accounts accept keys until
AllowSharedKeyAccessis false14. Google Cloud blocks service account keys in organizations created since May 202421 and disables exposed ones5. - Instance metadata. Require IMDSv2, which uses token-backed sessions, by account default or declarative policy. Track the
ImdsV1EnabledandMetadataNoTokenmetrics to zero, then addec2:RoleDeliveryto an SCP so IMDSv1 credentials stop working18. - Broad roles for workloads. Remove wildcard actions (Security Hub controls IAM.1 and IAM.21)15 and pin GitHub OIDC trust to one repository and branch or environment20. An over-broad deploy role took UNC6426 from a stolen GitHub token to AWS administrator in under 72 hours5.
- Guests and external access. Let only admins and Guest Inviters invite guests, and only from approved domains2; the external sharing sheet explains how this now also limits SharePoint sharing. Allow Teams external domains one by one and block chats from unmanaged accounts. Google Cloud's
iam.allowedPolicyMemberDomainsdoes the same for IAM grants21. SharePoint links: see the DLP sheet. - OAuth app consent. Microsoft's managed default for new tenants still lets users consent to permissions outside a list of mail, file, calendar and chat scopes10. Changes keep existing grants9, so review those too. Stolen Salesloft Drift tokens let attackers bulk-export Salesforce data5.
What changed in 2026
- Exploits overtook weak credentials in Google Cloud. In late 2025, software flaws opened 44.5% of Google Cloud intrusions in Google's data and weak or missing credentials 27.2%; misconfiguration fell to 21%. Google suggests its secure defaults may have driven the shift5. Patch within the patching sheet's deadlines.
- ScubaGear 2.0 is out. The October 2026 release adds a Security Suite baseline for Defender, Purview, Exchange Online and Teams3, so replace
defenderwithsecuritysuitein your config1. The Entra baseline now requires blocking risky AI agents where Agent 365 or E7 is licensed2. - Microsoft 365 moved toward secure defaults. Baseline Security Mode, in the admin center since November 2025, covers phishing-resistant MFA for admins, legacy authentication and app consent8. Tenant Configuration Management, generally available since April 2026, monitors over 200 settings in six workloads11.
- S3 public access can be blocked for a whole organization. Since November 2025 one AWS Organizations policy blocks it in every account, including new ones16.
What you can skip
- A paid CSPM before the free checks. ScubaGear and Defender for Cloud's Foundational CSPM cost nothing112. Buy attack path analysis once the basics are closed.
- Perfecting 90-day key rotation. Security Hub checks for it15, but AWS advises not creating long-term keys at all17. Remove them instead.
- A home-grown baseline. Start from a published one that someone keeps current.
Mistakes that cost the most
- Trusting a new default to fix old resources. Enforcing IMDSv2 for an account leaves instances that already allow IMDSv1 alone18, and consent changes keep old grants9.
- Running workloads in the AWS management account. SCPs and RCPs don't restrict it19. Keep it for billing and organization tasks.
- Exposing a test resource "just for a day." Microsoft measured a median of 5.3 hours from an exposed container's start to the first exploit attempt7.
What this looks like at your size
Small team, no security staff
- Turn on each Baseline Security Mode setting with zero impact8.
- Ask your IT provider for a monthly ScubaGear report and the list of consented apps and guests.
- Block public storage for the whole cloud organization, not account by account.
Growing, with IT or security staff
- Run ScubaGear weekly from a scheduled job, guard its app certificate like an admin credential and post the diff to the team channel.
- Connect every cloud account to one posture tool and give each finding an owner and a due date.
- Make infrastructure as code the only write path, and keep console write access for emergency roles.
Enterprise or regulated
- Treat BOD 25-01's mandatory SCuBA policies as your floor, as US federal civilian agencies must4.
- Feed drift and posture findings into the SIEM, with a fix time per severity.
- Review changes to guardrail policies like Tier 0 code, with two people.