Object storage has no application in front of it, so anyone holding a valid key, a signed URL or a public link can read the data straight through the cloud API, and by default nobody logs the read. MITRE ATT&CK T1530 Data from Cloud Storage covers adversaries collecting data from IaaS object stores such as Amazon S3, Azure Storage and Google Cloud Storage, and from SaaS storage such as OneDrive and Google Drive. In ATT&CK v20 it sits in the Collection tactic (TA0009) and applies to IaaS, SaaS and Office Suite platforms. This guide covers how the technique works from a defender's point of view, the data-plane logging you must switch on before you can see it, detection logic, and the provider settings that prevent it. It is for cloud security engineers, detection engineers and storage owners. Signed URL and SAS specifics are in the Azure SAS Tokens Security Guide, which this page does not repeat.
Key takeaways
- Two roads in. ATT&CK names misconfigured public or over-broad access and leaked credentials from repositories or logs; both end in ordinary API reads.
- Object reads are not logged by default. CloudTrail does not log S3 data events, Google Cloud Data Access logs are off, and Azure Storage resource logs need a diagnostic setting.
- The identity is the signal. A new principal, a key used from a new network or a sudden spike in
GetObjectvolume is what separates theft from normal use. - Public access controls exist at every level. S3 Block Public Access, Azure
AllowBlobPublicAccessand Google Cloud public access prevention stop the misconfiguration road. - Remove long-lived keys. Static access keys and storage account keys are the commonest leaked credential behind bucket theft; role-based, short-lived credentials shrink the window.
How the attack works
ATT&CK notes that IaaS storage usually has no overarching application such as a database engine, so data is retrieved directly through the Cloud API (T1059.009). SaaS storage can likewise be read through APIs rather than the front end, which is the boundary with T1213 Data from Information Repositories.
| Access route | What makes it possible | What defenders see |
|---|---|---|
| Anonymous access to a public bucket or container | Public ACL, public bucket policy, container access level set to Blob or Container, anonymous sharing links | Requests with no principal (Azure AuthenticationType "Anonymous"); nothing at all if data logging is off |
| Leaked long-lived key | AWS access keys, Azure storage account keys, GCS HMAC keys found in code, front-end files, CI logs or images | A known principal or key used from a new IP, ASN or user agent |
| Over-broad grant | "All authenticated users" style grants, wildcard principals, roles that include list and read on every bucket | Legitimate-looking calls from identities that have never touched the data |
| Signed URL or SAS | Long expiry, account-level scope, signed with an account key | Azure "SAS" authentication type, with no link to the person who created it |
| Stolen SaaS token or consented app | OAuth consent to an external application, session theft | Consent events followed by high-volume file downloads |
Collection is normally preceded by discovery of buckets and objects, covered in the T1580 and T1619 Cloud Infrastructure and Storage Discovery Guide, and often followed by T1537 Transfer Data to Cloud Account.
Where it shows up
- Extortion crews: ATT&CK records ShinyHunters collecting data from insecure cloud buckets, and Scattered Spider enumerating cloud resources for collection and searching OneDrive for VPN and MFA enrolment information and help desk instructions.
- Hybrid ransomware: Storm-0501 modified Azure Storage account resources through the
Microsoft.Storage/storageAccounts/writeoperation to expose accounts that were not remotely accessible, for exfiltration. For how the group first reached the cloud tenant, through stolen Entra Connect Sync credentials, see the Storm-0501 hybrid cloud attacks. - Espionage: HAFNIUM exfiltrated data from OneDrive; APT42 collected data from Microsoft 365; Fox Kitten obtained files from victims' cloud storage.
- Destructive operations: in the 2025 Poland wiper attacks, ATT&CK records stolen credentials used to download data from SharePoint and Teams.
Impact
Buckets often hold the data an attacker most wants and the defender least expects to be reachable: backups, exports, logs containing tokens, analytics extracts and customer documents. ATT&CK lists credit cards, personal data and medical records among the exposed data types. Theft through the storage API leaves no trace on endpoints or in the application, so without data-plane logging there may be no evidence of what was read, which matters for breach notification scoping.
Detection
Turn on the logs first
| Provider | Control-plane events (on by default) | Data-plane reads (off until enabled) | Notes |
|---|---|---|---|
| AWS S3 | CloudTrail management events: PutBucketPolicy, PutBucketAcl, PutPublicAccessBlock, DeleteBucket | CloudTrail data events: GetObject, ListObjects, ListObjectsV2, CopyObject, HeadObject | Data events cost extra; use advanced event selectors by bucket ARN and readOnly. Server access logs are an alternative, delivered on a best-effort basis |
| Azure Storage | Activity log: Microsoft.Storage/storageAccounts/write, listkeys/action | Resource logs via a diagnostic setting into the StorageBlobLogs table | Do not send logs to the account being monitored; set retention in the workspace |
| Google Cloud Storage | Admin Activity (ADMIN_WRITE): IAM policy changes, HMAC key creation | Data Access: ADMIN_READ (list buckets), DATA_READ (get and list objects) | Google states Data Access logs are disabled by default; Admin Activity cannot be disabled |
If full data logging is too costly, log reads on the buckets that hold regulated or secret-bearing data, and every access to buckets that should never be read by humans. See the Cloud Logging and Audit Trail Guide.
Detection logic
The ATT&CK v20 detection strategy for T1530 chains a spike in object access from a new IAM user or role with transfer to external addresses; for SaaS, an OAuth grant to an external app followed by high-volume downloads; and for Office, anonymous link creation followed by mass download.
- Volume anomaly per principal: objects or bytes read per identity per hour against a 30-day baseline; workloads are very regular, so deviations stand out. See the Service Account Behavioural Baselining Guide.
- Key used from a new network: any access key or service account key whose source IP or ASN has not been seen before, particularly hosting providers and Tor.
- Anonymous reads: any successful request with Azure
AuthenticationType == "Anonymous", or an S3 request with no principal, on a bucket not intended to be public. - Exposure changes: Block Public Access disabled, a bucket policy granting a wildcard principal,
AllowBlobPublicAccessset to true, or a container access level changed. - Managed findings: Amazon GuardDuty S3 Protection raises
Exfiltration:S3/AnomalousBehavior,Exfiltration:S3/MaliciousIPCaller,Policy:S3/BucketAnonymousAccessGranted,Policy:S3/BucketBlockPublicAccessDisabledandStealth:S3/ServerAccessLoggingDisabled. The policy findings come from management events; the rest need S3 data events. See the Amazon GuardDuty Guide.
Metrics
- Data-plane logging coverage: share of buckets classified sensitive with read logging enabled.
- Public exposure: number of buckets or containers reachable anonymously that are not on an approved public list, target zero.
- Static storage keys: count of active long-lived access keys, account keys and HMAC keys with storage permissions, and their median age.
- Time to revoke: minutes from an anomalous-read alert to key deactivation.
- Least-privilege drift: share of principals with read on all buckets in an account.
Prevention and hardening
- Block public access at the top (M1047 Audit, M1022): in AWS, enable all four Block Public Access settings (
BlockPublicAcls,IgnorePublicAcls,BlockPublicPolicy,RestrictPublicBuckets) at account or organisation level; S3 applies the most restrictive combination across levels. Note that removing a setting can make an existing public bucket accessible again. In Azure, setAllowBlobPublicAccessto false and enforce it with an Azure Policy Deny effect, after checking anonymous traffic inStorageBlobLogs. - Retire shared keys (M1018): in Azure, set
AllowSharedKeyAccessto false so only Microsoft Entra ID requests succeed; service and account SAS signed with the key are then rejected, while user delegation SAS still works. Use the built-in policy "Storage accounts should prevent shared key access" in Audit, then Deny. In AWS, replace user access keys with roles; see the AWS Access Keys Security Guide. - Least privilege on data roles (M1018): Microsoft notes that Owner, Contributor and Storage Account Contributor include
listkeys, which gives full data access, so restrict them to administrators and use Storage Blob Data roles for data users. - Network restriction (M1037): bucket policies conditioned on VPC endpoints or source IP, storage firewalls and private endpoints, so a leaked key does not work from the internet.
- Encrypt with customer-managed keys (M1041): ATT&CK advises including key rotation in the storage breach response plan; key policies add a second permission check on reads.
- Multi-factor for humans (M1032): require MFA for console and API access to sensitive storage where the platform supports it.
Choosing controls
| Situation | First steps | Trade-off |
|---|---|---|
| Few accounts, small team | Organisation-level public access block, GuardDuty or equivalent, remove user access keys | Managed detection has gaps without data events on key buckets |
| Large multi-account estate | Service control policies preventing public access changes, data events on classified buckets, central SIEM analytics | Data event cost; scope with selectors |
| Legitimate public content | Separate account or storage account for public assets, served through a CDN with origin access control | An extra account to manage, but sensitive buckets keep a strict block |
| Third parties needing data | Time-limited, object-scoped signed URLs or user delegation SAS, never account keys | More integration work for partners |
Phased rollout
- Inventory and classify: list every bucket and container, its owner, its data class and whether it is meant to be public.
- Observe: enable data-plane logging on sensitive stores and review anonymous and Shared Key traffic for two to four weeks.
- Audit-mode policy: assign public access and shared key policies in Audit; fix or document each non-compliant resource.
- Enforce: switch to Deny and organisation-level blocks; rollback is an exception for a single named resource, not removing the block.
- Detect and test: deploy volume and new-network analytics, then run an authorised read with a test key from an unusual network to confirm the alert and revocation path.
Where identity fits
Almost every T1530 case without a public bucket is a non-human identity problem: an access key, storage account key, HMAC key or service account token with more reach than its job needs, stored somewhere it can leak. Inventory these keys with owners, scope each one to the buckets it serves and prefer workload identity federation. See the OWASP NHI Top 10 NHI5 Overprivileged NHI Guide and the Cloud Access Key Leak Playbook.
Lessons from incidents
- Front-end code is public: in the Football Australia AWS keys exposure, long-lived keys in a public web page reportedly reached 127 buckets, one of which needed no authentication at all.
- Repository secrets reach storage: in the Sisense breach, a credential in a self-managed GitLab repository reportedly gave access to S3 buckets from which terabytes of customer data were copied.
- Configuration changes are part of the attack: Storm-0501 changed storage account settings to make data reachable, which is why exposure-change alerts matter as much as read alerts.
Framework mapping
| Framework | References |
|---|---|
| ATT&CK mitigations | M1018 User Account Management, M1022 Restrict File and Directory Permissions, M1037 Filter Network Traffic, M1041 Encrypt Sensitive Information, M1047 Audit, M1032 Multi-factor Authentication |
| NIST SP 800-53 Rev. 5 | AC-3 Access Enforcement, AC-6 Least Privilege, AU-2 Event Logging, SC-28 Protection of Information at Rest, SI-4 System Monitoring |
| NIST CSF 2.0 | PR.DS-01 (data at rest protected), PR.AA-05 (access permissions managed with least privilege), DE.CM-03 (technology usage monitored) |
| CIS Controls v8.1 | Control 3 Data Protection, Control 6 Access Control Management, Control 8 Audit Log Management |
| MITRE D3FEND | D3-RAPA Resource Access Pattern Analysis, D3-CRO Credential Rotation, D3-NRAM Network Resource Access Mediation, D3-FE File Encryption |
Practitioner checklist
- Enable S3 Block Public Access at organisation or account level and audit every exception.
- Set
AllowBlobPublicAccessandAllowSharedKeyAccessto false on Azure storage accounts, enforced by policy. - Turn on data-plane logging (CloudTrail data events,
StorageBlobLogs, GCS Data Access) for sensitive buckets. - Alert on anonymous reads, exposure changes and per-principal read spikes.
- Remove long-lived user access keys and storage account keys; use roles and federation.
- Restrict sensitive buckets to private endpoints or VPC endpoint conditions.
- Scan code, front-end assets and CI logs for cloud keys, and revoke any found.
- Keep a revocation runbook that also reviews data-plane logs for what a leaked key read.
Standards and references
- MITRE ATT&CK: Data from Cloud Storage (T1530)
- AWS: Blocking public access to your Amazon S3 storage
- AWS: Amazon S3 CloudTrail events
- AWS: GuardDuty S3 Protection finding types
- Microsoft Learn: Remediate anonymous read access to blob data
- Microsoft Learn: Prevent Shared Key authorization for an Azure Storage account
- Microsoft Learn: Monitor Azure Blob Storage
- Google Cloud: Cloud Audit Logs with Cloud Storage
Related NHI Mgmt Group resources: Amazon S3 Security Guide · Azure Storage Security Guide · Google Cloud Storage Security Guide · ATT&CK Collection Tactic Guide · Cloud Storage Exposure Playbook