How to Detect Data Exfiltration From Cloud Accounts
Attackers no longer break in to steal cloud data. They log in. Learn how to detect data exfiltration when every action looks legitimate.
How to Detect Data Exfiltration From Legitimate Cloud Accounts
The attacker in a modern data theft rarely breaks in. They log in. A stolen session token, a phished password or an over-privileged service key gives them a valid identity, and from that point every action they take looks like work: opening files, running exports, creating share links, syncing folders.
That is why so many organisations only learn about cloud data theft when the extortion email arrives. In his speech launching the NCSC Annual Review 2026, NCSC chief executive Richard Horne described the rise of "smash and grab" ransomware, where criminals steal data and threaten to publish it. The NCSC's own lessons from the year's incidents put it plainly: "people remain the target", and credential theft remains a common route in.
To detect data exfiltration in this situation, you cannot look for something that is blocked or malicious. You have to spot authorised actions that do not make sense. This guide explains how.
What Is Data Exfiltration From a Legitimate Cloud Account?
Data exfiltration is the unauthorised transfer of data out of an organisation's control. When it happens through a legitimate cloud account, the attacker or malicious insider uses a real, valid identity, such as a user, admin, service account, API key or connected app, to move data using the platform's own features.
No malware is needed, no firewall rule is broken and the traffic often never touches the corporate network. The data moves from one cloud service to another, authorised by a token the platform trusts.
Why Traditional Exfiltration Detection Misses It
Most data exfiltration detection was designed for a perimeter that no longer holds the data. Network egress monitoring watches traffic leaving the office or data centre, but cloud-to-cloud transfers never pass through it. Endpoint data loss prevention watches managed laptops, but a stolen token used from an attacker's machine never touches them. Signature-based tools look for known bad files, but an export of a customer table is a perfectly normal file.
The result is a detection gap shaped like trust. The more legitimate the account, the less visible the theft. As one recent analysis put it, identity is becoming a fundamental security barrier, and that applies to data protection as much as to access control.
Eight Ways Data Leaves Through Trusted Accounts
Before you can detect cloud data exfiltration, you need to know what it looks like on your platforms. These are the data exfiltration methods we most often see used with legitimate credentials.
| Method | What the attacker does | Signal to watch |
| --- | --- | --- |
| External sharing | Creates "anyone with the link" or external-guest shares on sensitive folders | New external shares on high-value data, especially in bulk |
| Storage permission change | Makes a storage bucket or container public, or adds an external principal | Policy or access-control changes on data stores |
| Snapshot or backup sharing | Shares a disk, database snapshot or backup with an external cloud account | Snapshot permission changes and cross-account copy events |
| Replication rules | Configures replication or sync to a destination the organisation does not own | New replication targets or sync destinations |
| Bulk export | Uses reporting, export or API functions to pull entire datasets | Export volume or row counts far above the identity's normal |
| Third-party app consent | Grants a malicious or unknown app broad read access through OAuth | New app consents with wide data scopes |
| Mail forwarding | Creates forwarding or inbox rules that copy mail to an external address | New forwarding rules, especially to personal domains |
| Sync to unmanaged devices | Signs in from an unmanaged device and syncs entire libraries | Large sync activity from new or unmanaged devices |
Notice that every signal on the right comes from the cloud platform's own audit and activity logs, not from the network. That is where cloud data exfiltration monitoring has to live.
The Five-Question Test for Unusual Data Access
Legitimate-account exfiltration does not break rules. It bends patterns. The most reliable way to catch it is to test significant data activity against five questions. One unusual answer is often noise. Two or three together are worth an analyst's time.
1. Who is acting, and is this normal for them? A finance analyst exporting finance data is normal. The same analyst exporting HR records is not. Non-human identities matter just as much: a service account that suddenly reads data it has never touched should be treated as a priority alert. Over-permissioned identities make this harder, which is why we focus on cloud identity risk from excessive permissions as a detection issue, not just an access issue.
2. What data is being touched? Volume alone misleads. Ten megabytes of board papers matters more than ten gigabytes of marketing images. Detection needs to know where the sensitive data lives and weight alerts accordingly.
3. Where is it going? Any destination outside the organisation's tenancy, such as an external account, personal email domain, unknown app or unmanaged device, raises the risk sharply.
4. How is it being moved? Methods matter. A user downloading files in a browser behaves differently from an API token paging through every record, or a new replication rule copying an entire store overnight.
5. When, and in what context? Activity at 3am from a new country, straight after a password reset or MFA change, during a resignation notice period or immediately after an impossible-travel sign-in is far more suspicious than the same action on a normal working day.
The five questions also give analysts a shared language for triage. An alert that says "unusual volume" is easy to dismiss. An alert that says "service account, HR data, external destination, API export, first time ever" is not.
Watch the Sequence, Not the Single Event
Exfiltration through a valid account is rarely one action. It is a short story with recognisable chapters, and the early chapters are often easier to detect than the theft itself.
- Secure the foothold. A new MFA method is registered, a new API key or app password is created, or session tokens are refreshed from an unfamiliar location.
- Explore. The identity browses or searches widely across sites, buckets or databases it rarely uses.
- Prepare. Permissions are widened, sharing settings changed, or data is gathered into one location or archive.
- Move. Exports, shares, syncs or replication carry the data out.
- Cover. Share links are deleted, forwarding rules removed, or audit settings changed.
Alerting on steps one to three gives the security team a chance to act before the data leaves. Correlating them across a few hours or days turns weak individual signals into a strong case. This is where a SOC that can join identity, cloud and data events together has a clear advantage over standalone tools.
How to Detect Data Exfiltration: A Practical Roadmap
Step 1: Switch on the logs that matter
Many cloud services do not record data-level activity, such as individual object reads, by default, or keep it for only a short period. The NCSC's lessons from the year's incidents list logging and monitoring among the factors that decide how an incident turns out. Check that data access, sharing, permission and admin events are logged across your main cloud and SaaS platforms, retained long enough to investigate, and sent to a central place where they can be correlated.
Step 2: Map your crown-jewel data
Identify the handful of data stores whose theft would cause real harm: customer records, financial data, intellectual property, HR files, legal documents. Detection effort should be concentrated there first.
Step 3: Baseline every identity, including the non-human ones
Build a picture of normal for each user and service identity: what data it accesses, how much, from where and when. Service accounts and API keys are often the most predictable identities and therefore the easiest to baseline, which makes deviations stand out clearly. Our guide to service account security covers why these identities are so often abused.
Step 4: Alert on control changes, not just data movement
Permission changes, new external shares, new forwarding rules, new app consents and new replication targets are low in volume and high in value. They make excellent detection rules because legitimate changes are relatively rare and easy to verify.
Step 5: Plant decoys where attackers will look
Canary files, decoy records and honeytoken credentials placed in sensitive locations have no legitimate use. Any access to them is a high-confidence alert, with almost no false positives.
Step 6: Tie identity risk to data activity
A risky sign-in on its own may be dismissed. A risky sign-in followed by bulk access to sensitive data should not be. Correlating identity signals, such as MFA phishing and session hijacking indicators, with data events is what turns a suspicious login into a confirmed incident.
Step 7: Separate insider risk from account takeover
The same signals can mean an external attacker using stolen credentials or an insider taking data before leaving. The detection logic is similar, but the response is very different. Behavioural analytics for insider threat detection helps distinguish the two by comparing activity with an individual's established patterns and role.
What to Do in the First Hour
When an alert points to active exfiltration through a legitimate account, speed and completeness both matter.
- Revoke sessions and tokens, not just the password. Resetting a password does not always end existing sessions, refresh tokens, API keys or app consents. Remove them all.
- Remove what the attacker created. Delete new shares, forwarding rules, replication targets, app consents and access keys.
- Preserve the evidence. Export the relevant audit logs before retention windows expire.
- Scope what left. Establish which data was accessed and transferred, and to where. This drives every later decision.
- Start the regulatory clock check. If personal data is involved, UK GDPR generally requires notifying the Information Commissioner's Office within 72 hours of becoming aware of a reportable breach.
These steps belong in a pre-agreed playbook, so that nobody is working out token revocation for the first time during a live incident.
Is Your Organisation Ready? Five Questions to Ask This Week
- If a valid account exported our most sensitive dataset tonight, would anyone be alerted, and how quickly?
- Are data access and sharing events logged on every major cloud and SaaS platform we use?
- Do we know which service accounts and API keys can read our crown-jewel data?
- Would we notice a new external share, forwarding rule or app consent on a sensitive system?
- Can we revoke every session, token and key for a compromised identity within minutes?
If more than one answer is uncertain, legitimate-account exfiltration is a live blind spot.
Turn Trusted Access Into Visible Access
Attackers have learned that the quietest way to steal cloud data is to use the access an organisation has already granted. Detecting it means watching how identities use data, not just whether they are allowed to.
Cyber Advisory Services helps organisations close this gap, from finding the logging gaps, over-permissioned identities and risky sharing settings attackers rely on, to correlating identity, cloud and data activity around the clock through our managed detection and response service. Speak with our team to find out whether your cloud accounts could be used to move data out without anyone noticing.