Architecture diagram
Three CASB deployment modes on one scene
A managed laptop steers cloud traffic to a forward proxy, an unmanaged device reaches a sanctioned application through a reverse proxy after single sign-on, and an out-of-band connector reads the same tenant through its API. One data policy serves all three, but only the inline modes can stop anything before it happens.
Step 1 / 5Get into the path, or see nothing
Architectural planes
Annotations
The same architecture, read top to bottom
Managed endpoints
MANAGEDOperated or administered on the enterprise's behalfSalesperson · Managed laptop · Insider or hijacked account
- PersonSalespersonWorking through a notice period
- DeviceManaged laptopAgent steers cloud traffic to the proxy
- AdversaryInsider or hijacked account
Reaches this boundary
- BlockedInsider or hijacked account stopped at Forward-proxy positionRefused: customer records to an unsanctioned service
Within this boundary
- RequestSalesperson to Managed laptopExports a customer list from the CRM
Unmanaged and personal devices
UNTRUSTEDNot controlled and not trusted; anyone may be present on itPersonal device
- DevicePersonal deviceNo agent, so nothing to steer traffic
Cloud access security broker
MANAGEDOperated or administered on the enterprise's behalfData policy
- Policy decisionData policyAllow, coach, redact, quarantine or block
Inline positions
MANAGEDOperated or administered on the enterprise's behalfForward-proxy position · Reverse-proxy position
- Security controlForward-proxy positionInline, and only if traffic is steered to it
- Security controlReverse-proxy positionInline via sign-on, no agent required
Out-of-band positions
MANAGEDOperated or administered on the enterprise's behalfOut-of-band connector · Shadow-IT discovery
- API or gatewayOut-of-band connectorReads the tenant after the event
- Security controlShadow-IT discovery
- Service catalogue with risk ratings
- Users and volume per service
- Newly seen services
- Sanction, restrict or block advice
Reaches this boundary
- RequestManaged laptop to Forward-proxy positionAgent steers cloud traffic to the proxy
- RequestIdentity provider to Reverse-proxy positionSession redirected through the broker after sign-on
- RequestPersonal device to Reverse-proxy positionUnmanaged device, covered without an agent
- TelemetrySanctioned CRM tenant to Out-of-band connectorActivity, file and sharing state read through the API
- TelemetryFirewall and web proxy logs to Shadow-IT discoveryLogs ingested to build the service catalogue
- AttackInsider or hijacked account to Forward-proxy positionCustomer list uploaded to a personal account
Within this boundary
- RequestForward-proxy position to Data policyUpload presented for a verdict before it leaves
- RequestReverse-proxy position to Data policySession and download presented for a verdict
- ContextOut-of-band connector to Data policyFindings evaluated against the same policy
- ContextShadow-IT discovery to Data policyService risk rating and usage volume
Enterprise control plane
INTERNALOperated by the enterprise itselfIdentity provider · Security operations · Firewall and web proxy logs
- Identity providerIdentity providerSingle sign-on, and the reverse-proxy hook
- Security controlSecurity operationsInvestigation and audit evidence
- Data storeFirewall and web proxy logs
Reaches this boundary
- ContextSalesperson to Identity providerSingle sign-on to the sanctioned tenant
- TelemetryManaged laptop to Firewall and web proxy logsOutbound destinations recorded at the network edge
- TelemetryShadow-IT discovery to Security operationsNewly seen services and who is using them
- TelemetryData policy to Security operationsEvery verdict and the signals behind it
Sanctioned cloud tenant
THIRD PARTYOperated by another organisation under its own termsSanctioned CRM tenant · Files and sharing links
- ApplicationSanctioned CRM tenantConnected by API and reachable inline
- Data storeFiles and sharing links
Reaches this boundary
- RequestForward-proxy position to Sanctioned CRM tenantPermitted traffic continues to the tenant
- RequestReverse-proxy position to Sanctioned CRM tenantPermitted session reaches the tenant
- RequestData policy to Files and sharing linksSharing link quarantined two days after the fact
Within this boundary
- RequestSanctioned CRM tenant to Files and sharing linksApplication stores files and creates sharing links
Discovered, unsanctioned services
UNTRUSTEDNot controlled and not trusted; anyone may be present on itPersonal file-sharing service
- ApplicationPersonal file-sharing serviceDiscovered, never connected
Reaches this boundary
- AttackInsider or hijacked account to Personal file-sharing serviceThe same upload from a device with no agent
Legend
Participants
- PersonSomeone making a request, and the identity they assertCircle with a person mark
- DeviceAn endpoint the request is made from, with its own postureRounded rectangle with an inset panel
- ApplicationAn application that serves or holds dataRounded rectangle with a panel stacked behind it
- Identity providerThe authority that asserts who is askingHexagon marked ID
- API or gatewayA network endpoint that terminates or forwards connectionsCapsule marked API
- Data storeWhere data restsCylinder
- Security controlA protective control, and the control plane's own observationRounded octagon with a check mark
- AdversaryA hostile party attempting somethingDiamond marked with an exclamation mark
- Policy decisionWhere policy is evaluated and one verdict is reachedDiamond with a split-path mark on its trailing point
Relationships
- RequestTraffic asking to reach a destinationSolid line. Filled arrowhead.
- TelemetryEvents reported onward for observation — observation, not controlDotted line. Small open arrowhead.
- ContextA signal that informs a decision without carrying the request itselfDash-dot line. Open chevron.
- AttackAn attempt made by an adversaryShort-dashed heavy line. Barbed open arrowhead.
- BlockedAn attempt terminated by the control the line starts fromHeavy line that stops short of its target. Perpendicular bar, and deliberately no arrowhead: the attempt did not arrive.
Boundaries
- Internal boundaryOperated by the enterprise itselfSolid outline. Badge reads INTERNAL.
- Managed boundaryOperated or administered on the enterprise's behalfDashed outline with a subtle wash. Badge reads MANAGED.
- Third party boundaryOperated by another organisation under its own termsDash-dot outline. Badge reads THIRD PARTY.
- Untrusted boundaryNot controlled and not trusted; anyone may be present on itDotted outline with a hatched leading corner. Badge reads UNTRUSTED.
Text alternative
The request, step by step
Get into the path, or see nothing
Inline coverage is a placement problem. A managed device is steered to a forward proxy by its agent; an unmanaged device is pulled through a reverse proxy by the identity provider at sign-on. A device that is neither steered nor redirected is not covered.
- Request: Salesperson to Managed laptop, “Exports a customer list from the CRM”
- Request: Managed laptop to Forward-proxy position, “Agent steers cloud traffic to the proxy”
- Context: Salesperson to Identity provider, “Single sign-on to the sanctioned tenant”
- Request: Identity provider to Reverse-proxy position, “Session redirected through the broker after sign-on”
- Request: Personal device to Reverse-proxy position, “Unmanaged device, covered without an agent”
Emphasised at this step: Forward-proxy position.
Decide before the data moves
Both inline positions present the action to one data policy and wait for the verdict, so an upload can be refused, coached or redacted before anything leaves. This is the only point in the diagram where prevention is possible.
- Request: Forward-proxy position to Data policy, “Upload presented for a verdict before it leaves”
- Request: Reverse-proxy position to Data policy, “Session and download presented for a verdict”
- Request: Forward-proxy position to Sanctioned CRM tenant, “Permitted traffic continues to the tenant”
- Request: Reverse-proxy position to Sanctioned CRM tenant, “Permitted session reaches the tenant”
Emphasised at this step: Data policy.
Read the tenant out of band
The connector authenticates to the service's own API and reads activity, files and sharing state. It sees things no proxy can, including files placed there by other routes, and it sees them only once they already exist.
- Request: Sanctioned CRM tenant to Files and sharing links, “Application stores files and creates sharing links”
- Telemetry: Sanctioned CRM tenant to Out-of-band connector, “Activity, file and sharing state read through the API”
- Context: Out-of-band connector to Data policy, “Findings evaluated against the same policy”
Emphasised at this step: Out-of-band connector.
Act after the fact
The same policy applies to a finding, but the only available actions are retrospective ones: unshare, quarantine, revoke, notify. Between the event and the scan that found it, the exposure was real and unmitigated.
- Request: Data policy to Files and sharing links, “Sharing link quarantined two days after the fact”
Emphasised at this step: Files and sharing links.
Discover what nobody sanctioned
Egress logs from the network name the services in use and who uses them, without any integration with those services. Discovery produces a catalogue and a recommendation, not a control, and its findings feed both the policy and the evidence trail.
- Telemetry: Managed laptop to Firewall and web proxy logs, “Outbound destinations recorded at the network edge”
- Telemetry: Firewall and web proxy logs to Shadow-IT discovery, “Logs ingested to build the service catalogue”
- Context: Shadow-IT discovery to Data policy, “Service risk rating and usage volume”
- Telemetry: Shadow-IT discovery to Security operations, “Newly seen services and who is using them”
- Telemetry: Data policy to Security operations, “Every verdict and the signals behind it”
Emphasised at this step: Shadow-IT discovery.
Attempts shown, and where each one stops
Insider or hijacked account attempts:
- Attack: Insider or hijacked account to Forward-proxy position, “Customer list uploaded to a personal account”
- Attack: Insider or hijacked account to Personal file-sharing service, “The same upload from a device with no agent”
- Terminated here by Forward-proxy position — “Refused: customer records to an unsanctioned service”. The line stops short and takes a bar, not an arrowhead.
Annotations
Why this line is observation, not a request path
Relationship — Telemetry: Sanctioned CRM tenant to Out-of-band connector, “Activity, file and sharing state read through the API”
No user traffic travels here. The connector polls or subscribes to the service's API and receives a record of things that already happened, so the relationship is one of reporting rather than carriage. Drawing it as a request path would imply the broker sits between the user and the tenant, which in this mode it does not.
An action that runs backwards into the tenant
Relationship — Request: Data policy to Files and sharing links, “Sharing link quarantined two days after the fact”
This is the one thing out-of-band mode can enforce, and it reaches the stored file rather than the person who moved it. It repairs an exposure that already existed for as long as the scan interval, which is why the interval is a security parameter and not an operational detail.
Which position could refuse, and why only that one
Relationship — Blocked: Insider or hijacked account stopped at Forward-proxy position, “Refused: customer records to an unsanctioned service”
The bar means the attempt ended before reaching its destination. It starts at the forward proxy because that position was already in the path with the upload held open. Neither the connector nor a reverse proxy for a different application was anywhere near this flow.
Why this attempt has no refusal
Relationship — Attack: Insider or hijacked account to Personal file-sharing service, “The same upload from a device with no agent”
Nothing is between this device and that service. No agent steers the traffic, the destination has no connector, and the identity provider is not involved because the service does not use it. Discovery will name the service afterwards, which is a report, not an intervention.
Why the inline positions are grouped
Boundary “Inline positions” (managed)
Everything inside this boundary can hold a request open and change its outcome. That single property, and not the vendor or the console, separates preventive coverage from reporting. Both positions inside it depend on a steering mechanism that can be absent, misconfigured or deliberately avoided.
No agent, but not free of conditions
Component “Reverse-proxy position”
This position needs the application to federate to your identity provider and needs sign-on to redirect through the broker, so it covers only applications you control the login for. It rewrites what the client sees, which is where pinned certificates, native clients and machine-to-machine integrations break.
Discovered is not the same as controlled
Boundary “Discovered, unsanctioned services” (untrusted)
Services in this boundary appear in reports with names, users and volumes. Nothing in the broker sits between a person and them until somebody either sanctions the service and connects it, or blocks the category at an inline position that is actually in the path.
This catalogue is built from people's records
Component “Shadow-IT discovery”
Its input is a log of where the workforce went. That makes it a monitoring capability with a lawful basis, a retention period, a restricted query surface and something the workforce should be told, settled before ingestion rather than after the first report is shared.
The short version
Three things to take away
-
The deployment mode is the capability
Out-of-band API integration, forward proxy and reverse proxy see different things at different times. Asking what a broker can do without naming the mode produces an answer that is true for one of them and false for the others.
-
Discovery and control are separate jobs
Building a catalogue of the cloud services people use comes from egress logs and needs no integration. Doing anything about a specific service needs either a connector to it or a position in front of it.
-
It sees what it is placed in front of
A broker covers the sanctioned services it is connected to and the traffic it is inline for. Everything else is a gap, and a discovery report naming a service is not the same as covering it.
Mechanics
How it works
A cloud access security broker sits between the people in an organisation and the cloud services they use, so that usage becomes visible and, sometimes, controllable. It exists because moving applications and data outside the organisation removes the vantage point security teams used to have. NIST SP 800-144 describes that displacement plainly, and a broker is one answer to it: not a replacement for the provider’s own controls, but a way to see and govern your side of the relationship.
Everything useful about a broker follows from where it is placed. Out-of-band integration authenticates to a service’s own interface and reads activity, file contents and sharing state. The picture is rich, covers files that arrived by routes no proxy saw, and arrives after the event, so the available actions are repairs. A forward proxy sits in the outbound path and can therefore refuse an upload before it completes, on the condition that something steers the traffic to it. A reverse proxy is reached through the identity provider at sign-on, so it covers unmanaged devices without an agent, on the condition that you control the login for the application. Three positions, three different answers to the same question.
Shadow-IT discovery is a fourth, separate job, and it is usually the first thing a broker is bought for. It needs no integration with anything: it reads the egress logs a network already produces and turns them into a catalogue of services, users and volumes with risk ratings attached. It is genuinely valuable, and it is a report. Naming a service does not put anything between a person and it. It is also a record of where people went, which makes it a workforce monitoring capability with all the obligations that implies.
The limits are worth stating as flatly as the capabilities. Out-of-band mode cannot prevent, only repair, and the scan interval is the length of the exposure. Inline modes break on the cases that resist interception: pinned certificates, native desktop and mobile clients, unusual protocols, unattended integrations. Every mode covers a named set of services, so a broker sees the sanctioned traffic it is placed in front of and not the whole estate. And no mode supplies the classification the data policy depends on, or the identity process that decides who should have had the export in the first place.
One concrete situation
A customer list leaves for a personal file-sharing account
End of a notice period, mid-afternoon
-
Context
A departing salesperson exports a customer list from the sanctioned CRM tenant, which is a permitted action for that role, and then tries to upload the file to a personal file-sharing account.
-
Decision
The data policy classifies the file as customer records, and the target service is on the discovered-but-unsanctioned list, so the verdict is to refuse the upload and tell the user why.
-
Enforcement
The forward proxy is in the path because the managed laptop steers cloud traffic to it, so the upload is stopped before any data leaves. From a personal device with no agent, the same upload would not pass through it at all.
-
Observation
The next out-of-band scan of the CRM tenant reports that an export link on the same records had been shared externally two days earlier, and quarantines it after the fact.
Scope, in both directions
What it covers, and what it does not
Both lists are authored and required. A control's limits are part of what it is, so the second panel is not a caveat appended to the first — it is the other half of the answer.
What it protects
- Visibility of which cloud services people actually use
- Data leaving through sanctioned services the broker can observe
- Sharing and configuration state inside connected tenants
- Sessions to sanctioned applications it sits inline for
- Files already at rest in connected services, retrospectively
- Evidence of cloud usage for audit and investigation
What it does not solve alone
- Real-time prevention in out-of-band mode: findings arrive afterwards
- Traffic using pinned certificates or non-browser clients
- Services it has no connector for and no position in front of
- Security of the provider's own platform and tenant separation
- Identity governance and the joiner, mover and leaver process
- Data copied to removable media or read on a personal device
- Classification decisions, which have to exist before any policy works
Frequently confused
Adjacent terms, and how they relate
| Term | Primary job | Relationship |
|---|---|---|
| Security service edge (SSE) | Cloud-delivered inspection and access control | The wider service a CASB is normally one capability inside. Buying SSE usually includes a broker; buying a broker does not give you the rest. |
| Secure access service edge (SASE) | Converged connectivity and cloud-delivered security | A CASB is one service in the chain a SASE architecture selects from, and the one that carries the SaaS and data half of the policy. |
| Zero trust network access (ZTNA) | Brokered access to individual private applications | The same brokering idea applied to private applications rather than cloud services. They sit side by side and neither covers the other's estate. |
| Data loss prevention | Detecting and stopping sensitive data in motion or at rest | The engine a CASB usually applies rather than a competing product. A broker supplies the position and the cloud context; the data policy and the classification behind it come from elsewhere. |
| Secure web gateway | Inspecting and filtering general outbound web traffic | Overlaps heavily with forward-proxy mode and is often the same appliance. A gateway reasons about sites and categories; a broker reasons about applications, tenants, actions and the data inside them. |
Explained in full
Side by side, in one shared diagram
Before procurement
What to settle first
Deliberately unnumbered: these are the questions to answer, not a sequence to work through in order.
-
Define sanctioned before buying anything
A broker enforces a list it does not write. Agree which services are approved, who approves them, what happens to the rest and how long a decision stands, or the tool will simply report.
-
Choose the mode per application
Map each service on the list to the mode that covers it, and record the services where no mode is available. That table is the honest statement of coverage, and it is the thing a shortlist should be judged against.
-
Test the protocol and certificate edges early
Native desktop and mobile clients, pinned certificates, non-standard ports and unattended integrations are where inline modes fail. Test them with real clients rather than accepting a support matrix.
-
Treat discovery data as employee monitoring
The catalogue is built from records of what people visited. Decide the lawful basis, the retention period, who may query it and what the workforce is told, before the first log is ingested.
-
Get classification working first
Every data verdict rests on knowing which files matter. Without agreed classification and a tested detection set, policy either blocks indiscriminately or passes everything, and both get switched off.
-
Decide the failure mode for inline modes
Establish whether traffic fails open or closed when the proxy is unavailable, who can invoke a bypass, how bypasses are recorded and how long one may last.
-
Check the out-of-band cadence
Ask what the scan interval, event lag and API rate limits actually are for your tenant sizes, because that interval is the window in which a retrospective control does nothing.
Sources
-
NIST SP 800-144
Guidelines on Security and Privacy in Public Cloud Computing
Sets out the loss of visibility and direct control that follows moving data and applications outside the organisation, which is the gap a broker is bought to narrow.
-
NCSC
Separates protection of data in transit, separation between customers and authenticated access to service interfaces, which is how the provider's responsibilities and the broker's divide.
-
NIST SP 1800-35
Implementing a Zero Trust Architecture
Reference implementations showing where an enforcement point can sit relative to a SaaS application, including identity-mediated redirection.
-
CISA v2.0
Its data and applications pillars describe inventory, classification and access control as prerequisites, none of which a broker supplies for you.