Join our Newsletter — 33% off our NHI Course
Home› Guides› Cloud DLP vs Endpoint DLP: Covering Data Where…
Guide Data Security

Cloud DLP vs Endpoint DLP: Covering Data Where It Actually Moves

PDF for newsletter subscribers · free to read online

← All guides
By NHI Mgmt Group, edited by Lalit Choda Updated 14 November 2026 14 min read
On this page

Most data loss programmes buy either cloud DLP or endpoint DLP and assume the other channel is covered, when in practice each one leaves a gap the size of the other's strengths. Cloud DLP inspects data in and on its way to cloud services, through API connections to SaaS and storage, inline proxies in a security service edge, or controls built into the SaaS platform itself. Endpoint DLP inspects data on the device through an operating system agent, controlling USB, print, clipboard, network shares and uploads from that machine. This guide compares the two by channel coverage, inspection mode (inline, API and agent), GenAI paste and browser channels, false positive management and architecture by device and data type. It is for data security leads, SSE and endpoint teams, and architects deciding which control owns which flow. The browser-only option is covered in Browser DLP vs Endpoint DLP, and posture scanning of data at rest in DSPM vs DLP; this page links rather than repeats them. Where platform examples are named they are listed alphabetically; order is not a ranking.

Key takeaways

  • The short answer is device versus destination. Endpoint DLP controls what leaves a device you manage, through any channel; cloud DLP controls what reaches or moves inside cloud services, from any device, including ones you do not own.
  • Cloud DLP is really three controls. API mode scans data already in SaaS and storage after the fact, inline mode inspects traffic in real time through a proxy, and SaaS-native DLP enforces inside the platform. Each has different latency, coverage and bypass routes.
  • Endpoint DLP alone fails on unmanaged devices and SaaS-to-SaaS flows. A contractor laptop, a phone, or an OAuth app pulling files between tenants never touches the agent.
  • Cloud DLP alone fails on local channels. USB, local print, Bluetooth, RDP copy and desktop apps that use their own protocols are invisible to an API connector and often to the proxy.
  • One classification, one incident queue. Running both is normal; running both with different sensitive information definitions doubles false positives and splits investigations.

The three forms of cloud DLP

  • API mode (out of band): the DLP service connects to a SaaS or storage platform through its API, scans files and messages at rest or shortly after they change, and applies remediation such as removing a public link, quarantining a file or applying a label. Examples, alphabetically: Amazon Macie discovers sensitive data in Amazon S3 using automated sampling or targeted jobs; Google Cloud Sensitive Data Protection profiles BigQuery, Cloud SQL and Cloud Storage; Microsoft Purview DLP can scope policies to non-Microsoft cloud apps through Microsoft Defender for Cloud Apps. Strength: no traffic routing, covers data whoever put it there. Weakness: acts after the data has arrived, and API rate limits create a lag.
  • Inline mode (proxy): traffic passes through a forward or reverse proxy, typically part of an SSE platform or a CASB session control, which inspects uploads, downloads and form posts in real time and can block, warn or encrypt. Microsoft Defender for Cloud Apps session policies, for example, route sessions through Conditional Access App Control to block or protect downloads and uploads, block copy or print, and require step-up authentication; each policy applies to a single host app. Strength: blocks before the data lands. Weakness: only sees traffic that is routed through it, and TLS inspection, certificate pinning and native apps create blind spots.
  • SaaS-native DLP: the SaaS platform enforces rules on its own content. Google Workspace DLP covers Drive, Gmail, Chat and Calendar, with actions to block sharing, warn users and audit, and can condition rules on Context-Aware Access. Microsoft Purview DLP covers Exchange, SharePoint, OneDrive, Teams and Microsoft 365 Copilot (preview). Strength: deep context such as sharing state and recipients. Weakness: one platform per policy engine.

For the CASB side of API and inline control, see the CASB Guide; for the endpoint agent in depth, the Endpoint DLP Guide.

Side-by-side comparison

DimensionCloud DLP (API)Cloud DLP (inline)Endpoint DLP
Where inspection runsProvider-side, via SaaS and storage APIsProxy in the traffic pathAgent on the operating system
TimingAfter data lands (near real time to hours)Real time, before data landsReal time, at the action
Device ownership neededNoNo for reverse proxy; agent or PAC for forward proxyYes in practice
Data at rest in SaaSYes, including historic dataNoNo
SaaS-to-SaaS and OAuth app flowsYes, when both ends are connectedNoNo
Local channels (USB, print, Bluetooth, RDP)NoNoYes
Typical remediationRemove sharing, quarantine, label, notify ownerBlock, warn, encrypt on downloadBlock, warn with override, audit
Main bypassUnconnected apps, API lagUnrouted traffic, pinned apps, personal networksUnmanaged devices, agent tamper, unsupported apps

Channel coverage gaps

Map channels, not products. The table shows which control can see each common exfiltration route; the gaps are where programmes fail.

ChannelATT&CKCloud APICloud inlineEndpoint
Upload to personal cloud storage from a managed laptopT1567.002Only if the destination tenant is connectedYes, if routedYes
External sharing link on a corporate SaaS fileT1567YesOnly at the moment of sharing, if routedNo
Download from SaaS to an unmanaged deviceT1530NoYes (session control)No
Copy to USBT1052.001NoNoYes
Transfer over a non-web protocolT1048NoPartial (depends on proxy scope)Partial (network share, restricted apps)
Email with attachmentT1048Yes (SaaS-native mail DLP)Webmail onlyPartial
Third-party app pulling files via OAuthT1530Yes (connected platform)NoNo
Paste into a GenAI siteT1567NoYes, with inspection of form postsYes on supported browsers

Two gaps recur. First, data in SaaS shared onwards by a user or an integration is invisible to the endpoint. Second, anything that leaves through a local channel is invisible to every cloud mode. A programme with only one of the two has to accept one of those gaps explicitly.

Option profiles: threat model, architecture, failure modes

Cloud DLP

  • Threat model: accidental oversharing in SaaS, public links, sensitive data uploaded to unsanctioned services, downloads to personal devices, and integrations or compromised accounts that pull data out through APIs.
  • Architecture: API connectors to each sanctioned SaaS and storage platform for at-rest scanning and sharing remediation; an inline proxy (forward for all web traffic, reverse for specific apps) for real-time control; identity provider integration so policies know user, group and device posture.
  • Failure modes: only some tenants connected; inline policies scoped to a handful of apps; TLS inspection exceptions that swallow the riskiest destinations; API scans running days behind; remediation that notifies but never removes the link.
  • Controls that matter: connect every sanctioned tenant, enforce session control on unmanaged devices for sensitive apps, block unsanctioned storage categories, and remediate sharing automatically for the highest sensitivity labels.

Endpoint DLP

  • Threat model: insiders and careless users moving data through local channels, departing staff copying files, and data reaching unsanctioned desktop apps.
  • Architecture: an agent onboarded through device management, evaluating file activity against policy. Microsoft Purview Endpoint DLP, for example, supports Windows 10 1809 and later, Windows 11, Windows Server 2019 and later, and the three latest macOS versions, and can audit or restrict copy to USB, network share, clipboard, print, unallowed Bluetooth apps and RDP (Windows), as well as uploads to restricted cloud domains. A device-scoped policy applies only when both the user and the device are in scope.
  • Failure modes: agent missing on a share of the fleet, servers left out (Endpoint DLP on Windows Server must be explicitly turned on), unsupported file types passing silently, image content ignored without OCR, and Linux or VDI estates uncovered.
  • Controls that matter: coverage reporting against the device inventory, restricted app and USB device groups, block with justification for high-sensitivity data, and tamper alerts.

Inline vs API vs agent: choosing the inspection point

  • Choose API when the question is "what sensitive data is already in this SaaS, and who can see it?" or when the flow is SaaS to SaaS. It is the only mode that sees historic content and integration traffic.
  • Choose inline when you must stop data before it reaches a destination, especially from devices you cannot install an agent on, and when the destination is a web application.
  • Choose agent when the channel is local, the application is a desktop app, or you need control regardless of network path, including offline. Microsoft documents that existing policies still apply to already-classified files when a device is offline.
  • Combine for high-sensitivity data: SaaS-native or API controls on the repository, inline session control for unmanaged access, and the agent on managed devices.

GenAI paste and browser channels

Pasting into AI assistants is where the three modes overlap most and where responsibilities blur. Endpoint agents can evaluate paste to supported browsers against restricted service domains; inline proxies can inspect form posts to AI sites if traffic is routed and decrypted; API mode does nothing for consumer AI services, though SaaS-native controls increasingly cover enterprise assistants. Microsoft Purview, for example, lists Microsoft 365 Copilot as a DLP location (preview) and offers inline protection for named AI sites through Edge for Business and through network integration with SASE products. Google Workspace DLP extends to Gemini and to Chrome through Chrome Enterprise Premium.

Decide once which control owns each AI channel: SaaS-native for sanctioned enterprise assistants, browser or inline for web AI from any device, endpoint for desktop AI apps and coding assistants on managed devices. Agent and MCP traffic is a separate problem; see GenAI DLP vs MCP DLP and GenAI DLP vs Traditional DLP.

False positive management across both

False positives drive users to workarounds and analysts to ignore queues. Running two DLP layers doubles the risk unless they share definitions.

  1. One classification source: use the same sensitive information types, custom detectors and sensitivity labels for cloud and endpoint policies, so one tuning change fixes both. Label-based rules generate far fewer false positives than raw pattern rules. See the Data Classification Guide.
  2. Simulate first: Microsoft Purview recommends draft, simulation, tune, then enable; simulation evaluates policies and reports matches without applying actions. Google Workspace rules can start with audit actions before warn or block.
  3. Use allow lists and thresholds: Amazon Macie supports allow lists for text or patterns to ignore, such as test data; most engines support instance counts and confidence levels. A rule on one card number generates noise; a rule on fifty does not.
  4. Warn with justification before block: overrides with a business reason give tuning data. Review override reasons monthly and tighten rules where justifications are weak.
  5. Read match context: use the snippet or contextual summary features (Purview Activity explorer, Workspace DLP snippets) to see why a rule fired before changing it.

Architecture by device and data type

Device or dataPrimary controlSupporting control
Corporate managed laptop, any dataEndpoint DLPInline for web, API on the SaaS repository
BYOD or contractor deviceInline session control (block download, protect)Browser DLP in a managed work profile
Mobile phoneInline session control and app protection policySaaS-native sharing rules
Structured data in cloud warehouses and bucketsCloud data discovery (Macie, Sensitive Data Protection)DSPM for access paths
Documents in collaboration SaaSSaaS-native DLP and API sharing remediationEndpoint on download to managed devices
Source codeCode host controls and secret scanningEndpoint restricted apps, inline for unsanctioned repositories
Servers and VDIEndpoint DLP where supportedNetwork egress control

Decision guidance

  • Mostly managed Windows and macOS fleet, mostly one SaaS suite: start with the suite's native DLP plus endpoint DLP on the same classification; add inline for unsanctioned web later.
  • Large contractor or BYOD population: inline session control and browser controls first; endpoint covers only the managed share.
  • Data concentrated in cloud data platforms: cloud discovery and DSPM first; DLP on the egress paths from those platforms.
  • Small organisation: use what the SaaS suite includes, enforce download restrictions for unmanaged devices, and add endpoint DLP only for teams handling regulated data.
  • Privacy constraints: endpoint agents see whole device activity. The ICO expects a DPIA where monitoring is likely to be high risk, transparency to workers and the least intrusive means; inline controls scoped to work apps are usually less intrusive on personal devices.

Phased rollout

  1. Inventory channels and data: list sanctioned SaaS tenants, storage, device populations (managed, BYOD, contractor, server) and the top sensitive data types.
  2. Agree one classification: sensitive information types and labels used by every layer.
  3. Connect APIs and audit: connect sanctioned tenants and run discovery; fix public and external sharing of the most sensitive data.
  4. Endpoint in audit: onboard devices in pilot rings, audit only, and check coverage against the device inventory.
  5. Inline for unmanaged access: session control on sensitive apps in report-only mode, then block downloads from unmanaged devices.
  6. Move to warn, then block: per channel, starting with the clearest cases (USB, public links, personal storage) for the highest labels.
  7. Rollback: each policy returns to simulation or audit if business impact appears; keep the change record.

Where identity fits

Cloud DLP decisions depend on identity: which tenant, which account, whether the device is compliant. They also depend on non-human identities: OAuth apps and service accounts with broad file read scopes can move data between services with no device involved, which only API-mode controls and app governance see. Review third-party app grants on SaaS platforms alongside DLP policy. See the CASB Guide and DSPM vs DLP for the access side.

Detection and measurement

  • Channel coverage: share of identified exfiltration channels with an enforcing control, from the channel table above.
  • Agent coverage: onboarded devices as a percentage of the managed inventory, by operating system.
  • Tenant coverage: sanctioned SaaS and storage tenants connected for API scanning.
  • False positive rate: share of alerts closed as false positive, per rule, trending down after tuning.
  • Override rate and quality: overrides per thousand matches, with share of justifications judged valid on review.
  • Exposure dwell time: time from a sensitive file becoming externally shared to remediation.

Common failure patterns

  • Assuming endpoint DLP covers SaaS: data shared onwards inside SaaS or pulled by an integration never touches the agent.
  • Inline policy on three apps: the proxy blocks uploads to the named services while users move to the next one.
  • Different definitions per layer: the same document is sensitive to one control and not the other, so incidents do not correlate.
  • Servers and VDI left out: bulk data often sits there, and server support may need explicit enablement.
  • Block on day one: users find workarounds, executives request exceptions, and the programme loses support.
  • No owner for API findings: thousands of overshared files are reported and none are fixed.

Framework mapping

FrameworkCloud DLPEndpoint DLP
NIST CSF 2.0PR.DS-01 data at rest, PR.DS-02 data in transit, ID.AM-07 data inventoriesPR.DS-10 data in use, PR.DS-01
NIST SP 800-53 Rev 5AC-4 information flow enforcement, SC-7 boundary protectionMP-7 media use, AC-4
MITRE ATT&CKT1530, T1567, T1567.002T1052.001, T1048, T1567
NIST SP 1800-28Data confidentiality: identifying and protecting assets against data breaches

Practitioner checklist

  • Build the channel table for your estate and name the control that owns each row.
  • Use one set of sensitive information types and labels for every DLP layer.
  • Connect every sanctioned SaaS and storage tenant for API scanning and sharing remediation.
  • Enforce session control that blocks or protects downloads on unmanaged devices for sensitive apps.
  • Onboard endpoint DLP across Windows, macOS and in-scope servers; report coverage monthly.
  • Assign each AI channel to one control: SaaS-native, inline or browser, or endpoint.
  • Run every new policy in simulation or audit, then warn, then block.
  • Review OAuth app grants with file scopes alongside DLP findings.
  • Complete a DPIA before endpoint monitoring where risk to workers is likely to be high.

Standards and references

Related NHI Mgmt Group resources: Data Loss Prevention (DLP) Guide, DLP Buyer's Guide, BYOD Security Programme, Insider Data Theft Guide.

PDF for subscribers

Take this guide with you

Every section, table and checklist in one PDF, with links to the sources and related guides. Subscribe to our weekly NHI & AI Identity Journal and we’ll email you a download link.

Free weekly newsletter

Subscribe to the NHI & AI Identity Journal

The latest on NHI and Agentic AI security – articles, research, breaches, news and events every week.

Bonus 33% off our NHI Course when you subscribe.

NHIMG Editorial Note
Researched and written by the NHI Mgmt Group editorial team using AI tools, under the editorial direction of Lalit Choda. Facts are drawn from the sources listed on this page and checked by our automated quality process. Last updated 14 November 2026. Spotted an error? Use our contact page.
NHI Mgmt Group, the independent authority on Non-Human Identity, IAM and Agentic AI security. nhimg.org

PDF Downloads for Newsletter Subscribers

Get the guide

Enter your details and we’ll email you a download link.

The full guide is free to read online. The PDF is a thank you for subscribing.

Fill in all fields marked * and tick the PDF box to continue.