Conditional access

The credential is right. Is the device?

Conditional access asks the second question. Before an application opens, the device has to answer for itself, and the answer is yes or no.

  • Threeways to establish the device, used together or alone
  • Anyapplication that speaks SAML, whichever suite it belongs to
  • Beforethe session exists, not after it has opened
What a sign-in proves

Three things a valid credential cannot tell you

A credential proves a secret was known

That is all it proves. It says nothing about the machine the secret was typed on, who else uses that machine, whether anyone has ever seen it before, or whether its protection was switched off last Tuesday.

Mail is the whole estate

Once a mailbox opens, so do the archive, the shared drive, the chat history and every attachment anyone ever sent. A personal machine with a valid credential holds all of it, and nobody manages that machine.

Second factors answer the wrong question

A one-time code confirms the person, which is worth confirming and is not the hard part. A phished credential with a relayed code still arrives from a machine nobody has approved, and nothing in the exchange notices.

What it is

A rule with an if in it

If the device is one you recognise, approved, and in the state your policy requires, let the request through. If it is not, do not, regardless of how valid the credential and the second factor were. The hard part was never the rule. It is that the rule has to be enforced where the application decides whether to trust the login, which is the identity provider.

The condition is evaluated at the identity provider

The application hands the sign-in over and InstaSafe decides, so the condition is settled before the application opens. Nothing is installed inside the application, and there is no per-application integration to maintain.

The answer is binary

The device satisfies the named rules or it does not. There is no score, no weighting and no judgement call, which is what makes the policy something an auditor can read rather than something an engineer has to explain.

The device has to be established first

A rule about a device needs a device it recognises. That recognition comes from binding, from management enrolment, or from an agent reading the machine's real state. Which one depends on whose hardware it is.

Every decision is written down

Each evaluation is recorded as an event against the connection, naming the condition that decided it rather than a bare allow or deny, and exports to a SIEM.

A tenant, and one way into it

Everything your people signed in to this morning

The applications the company runs, the mailboxes, the shared drives and the five people currently holding sessions. Keep scrolling, or give the glass a click.

One credential opened all of it.

Somebody signed in from a machine nobody had approved. The credential was valid, so the suite let it through, and everything above came with it: the mailboxes, the shared drives, the chat history, the admin centre. Nothing asked what the machine was, who manages it, or whether anyone had seen it before.

Establishing the device

A rule about a device needs a device it recognises

There are three ways to get one, and which you use is decided by whose hardware it is rather than by preference. Most estates need more than one: a managed laptop and a contractor's own machine cannot be established the same way, and refusing the contractor outright is how a policy gets switched off by a frustrated executive.

Device binding

Hardware you can enumerate but do not manage with a platform.

A device registers itself on first use and waits. Until an administrator approves it in the console, somebody signing in from it gets nowhere. Approved devices are bound to the people allowed to use them, so an approved machine is not a free pass for everyone in the directory.

  • A device nobody approved cannot be used
  • An attempt from an unapproved machine raises a notification
  • Approval is per person, not per machine
  • Devices that are lost or retired are revoked in one place

Binding answers whether a device is yours. It does not answer whether the device is healthy. Pair it with posture if you need both.

MDM enrolment

An estate already enrolled in mobile device management.

The management platform holds enrolment and compliance state, and that state becomes a condition like any other. Enrolment is a first-class check in the policy language, sitting alongside domain membership, so nothing has to be duplicated from the management console.

  • Compliance comes from the platform already deciding it
  • No second set of device policies to maintain
  • The access app is distributed through the management catalogue
  • Replacing the platform later does not mean rewriting access policy

On this path the access app is issued by the catalogue with a key, and a copy installed from a public store does not carry it. That property belongs to this path and not to the others.

Agent posture

Machines where you need the state of the device, not just its identity.

An installed agent reads what is actually true about the machine and hands it to the policy: protection present and running, definitions current, disk encrypted, operating system at or above a floor, required processes there and forbidden ones absent.

  • Conditions about health, not only identity
  • Checks are named facts, never inferred
  • Re-read when the connection is established
  • The same vocabulary across the estate

Posture needs the agent. Devices with nothing installed can still be conditioned, but on the gateway parameters below rather than on their internal state.

MDM path

Mobile device management: same app, two different answers

This section is the MDM path and nothing else. Where mobile device management distributes the access app, the copy the catalogue issues carries a key the public store copy does not. Flip the install source: everything about the two requests is identical except where the app came from, which is all the policy is asking about.

  • 01/04App sourceIssued by the management catalogue
  • 02/04Enrolment keyPresent and matched
  • 03/04ManagementActive on the device
  • 04/04AssertionIssued

AllowMicrosoft 365 session opens

The user did everything right except the part that mattered. Installing the app is not the same as being given it.

A request arrives at the login page. Nothing about it is known yet.

Unidentified session

The directory answers.

Arjun MActive Directory

A second factor is asked for, and checked against the record the directory already holds.

Authenticator app

Both factors hold.

Verified

The session is handed to the gateway that fronts the tenant.

gw-mumbai-01

The gateway loads what this person's groups are allowed to reach.

Partners and contractorsMail · Chat · Files

Everything that entitlement opens.

Microsoft 365TeamsJiraSalesforce

One of them is asked for.

Microsoft 365

The machine it came from carries no management agent, which is a hard condition, so nothing below it is evaluated. The event is written with the condition that decided it.

RefusedDESKTOP-4KQ2N1 · unenrolled
See the policy that decided it
Nothing on the device

Three conditions that need no agent and no enrolment

Auditors, a board member on a personal tablet, a contractor for one afternoon: the population where installing anything is unrealistic. These are read at the gateway as the sign-in arrives. They are not device posture, which needs the agent, and the two are worth keeping apart when you write the policy.

Known serial number

Access restricted to machines whose serial numbers are recorded. A serial that is not on the list does not proceed.

Trusted address range

Ranges the organisation defines. Requests from outside them do not proceed, whichever credential they carry.

Permitted location

Countries or cities from which access is allowed. A flat allow-list, read once as the sign-in arrives.

Fewer checks are available this way than with the agent. That is the honest trade-off for installing nothing, and it is usually the right one for people who will never be your employees.
What the gateway receives

A person sees a laptop. Policy sees the facts about it.

Drag the belt, or let it run. Everything to the right of the line is a machine somebody is sitting at. Everything to the left is what actually arrives.

10:39Sunday, 19 JulyPriya SCORPpriya.sPasswordWINDOWS SIGN-IN · INSTASAFE AGENT CONNECTED
10:39Sunday, 19 JulyAlen JosephCORPalen.josephPasswordWINDOWS SIGN-IN · INSTASAFE AGENT CONNECTED
10:39Sunday, 19 JulyNeha VCORP eha.vPasswordANDROID · WORK CONTAINER · AGENT CONNECTED
10:39Sunday, 19 JulySophia SCORPsophia.sPasswordWINDOWS SIGN-IN · NO FACTOR REGISTERED
10:39Sunday, 19 JulyArjun MWORKGROUParjunPasswordWINDOWS SIGN-IN · NO AGENT DETECTED
10:39Sunday, 19 JulyContract auditor(not joined)PasswordWINDOWS SIGN-IN · NO AGENT DETECTED
Where it applies

The mechanism has nothing suite-specific in it

Most of these conversations start with whichever suite holds the mail, because that is where the data is. What sits behind the decision changes nothing about how the decision is made.

Microsoft 365

Federate a domain and every product in the tenant inherits the condition, because they all inherit the sign-in. Outlook, Teams, SharePoint and the rest need no individual integration and none of them quietly stays outside it.

  • Federation is configured for a domain, not for a handful of people. That is Microsoft's rule, and it is why an evaluation starts on a secondary domain rather than the one payroll reads mail on.
  • Conditional access through Microsoft's own stack is gated behind higher licence tiers. Evaluating it at the identity provider instead does not depend on the tier.

Google Workspace

The same arrangement over SAML: the console becomes the identity provider for the domain, and sign-in is settled there before mail opens. Access can be refused for a person, a device, or a device in the wrong state.

  • Google Workspace does not offer SCIM provisioning, so people are onboarded in bulk rather than synchronised continuously. Worth knowing before a large rollout, because it decides who maintains the list.
  • Suspending somebody in the console refuses every new sign-in immediately. A browser session already open runs until it expires.

Anything else that speaks SAML

The mechanism has nothing suite-specific in it. Any application that accepts a SAML assertion can sit behind the same decision, alongside private applications reached over the gateway.

  • Support is a protocol question rather than a catalogue question: if it speaks SAML, it can be put behind this.
Underneath

Compliant is not a feeling. It is 144 named rules.

On the posture path the agent reads the device's actual state against 25 check types expressed as 144 named rules, across more than 1,500 operating-system and device combinations. A rule is a check plus a threshold plus who it applies to, which is what makes the policy something you can hand an auditor rather than describe.

The named check attributes

Session and identity

  • BrowserID
  • TenantId

Malware protection

  • Antivirus
  • AntiVirusStatus
  • AntiVirusLastUpdate
  • SEPLastAVUpdate
  • AntiSpyWare
  • AntiSpywareStatus

Operating system

  • OS
  • OSName
  • OSVersion
  • ServicePack
  • Hotfix

Running processes

  • ProcessRunning
  • ProcessNotRunning

Disk encryption

  • BitLocker

Domain and management

  • InDomain
  • DomainName
  • mdm

Network protection

  • Firewall
  • FirewallStatus

Files and registry

  • FileExists
  • FileNotExists
  • RegistryKeyExists

Worth noticing in that list: mdm is itself one of the named attributes. Management enrolment is not bolted onto the policy language from outside. It is a first-class condition in it, sitting alongside domain membership, which is why the three paths combine rather than compete.

Devices wait for a person

The first time a machine is seen it appears in the console as pending, with its details attached, and somebody signing in from it gets nowhere until an administrator approves it. An attempt from an unapproved machine raises a notification rather than failing quietly. Nothing self-enrols in the background.

  • Allow the first device, hold the rest for approval
  • Hold every new device for approval
  • Allow the first device from desktop and mobile, hold the rest
  • Approve new devices automatically after a set number of days
  • Approve nothing automatically

Approval is per person, not per machine

An approved device is not a free pass for everyone in the directory. Each person allowed to use a machine is bound to it separately, which is what lets a shared desk carry several people across several shifts while keeping their sessions apart. One person holds a session at a time.

The second factor is one condition among the set

It answers a question about the person, which is worth asking and is not the hard part. It can be enforced everywhere, or everywhere except a trusted address range, scoped org-wide, per group or per user. A phished credential with a relayed code still arrives from an unapproved machine, and that is what refuses it.

Where it earns its keep

Nobody has one kind of user. So nobody gets one rule.

Shift workers on shared machines

A floor of desktops used by different people across a morning and an evening shift. The machine is approved once; each person is bound to it separately, so the desk is shared and the sessions are not.

One machine, many approved people, one session at a time

Contractors on hardware you will never own

Partner engineers and contract staff who need mail and chat on their own machines. Domain membership is impossible, so it is traded for binding and posture rather than waived, and the access ends when the engagement does.

Weaker than a managed laptop, far stronger than a credential

Personal devices doing real work

A phone or a laptop somebody owns, carrying the work container. Access is permitted while the device meets the bar and refused the moment it stops, without the organisation taking ownership of the hardware.

Re-read when the connection is established

Somebody leaves on a Friday

Suspend the person and every new sign-in is refused at once, on every device they hold. Revoke the device and the same is true of the machine. Neither needs the hardware back.

New sign-ins refused immediately; open sessions run to expiry

What changes

Four things, one of which is an inventory you did not have

An unmanaged machine stops being a way in

The common intrusion is a valid credential on hardware nobody has approved. That path closes, and it closes for every application behind the same sign-in rather than one at a time.

The estate becomes something you can list

Every device that has ever attempted a sign-in is in one console with its details attached, which is usually the first accurate device inventory an organisation has had.

Policy survives a change of platform

Enrolment is a condition, not an integration. Replacing the management platform later does not mean rewriting access policy around whatever the new one calls things.

The decision is defensible

Each evaluation names the condition that decided it. A year later that is the difference between showing an auditor a record and describing an intention.

Getting there

One virtual machine, one population you can afford to break

  1. Stand up the gateway

    One virtual machine inside the network, on its own rather than sharing a host with existing web services. It needs to reach the directory, and that is the whole footprint.

  2. Bring the people in

    Users and groups come from the directory already in use. Where the directory supports continuous provisioning it synchronises; where it does not, people are onboarded in bulk and the list is maintained deliberately.

  3. Start on a population you can afford to break

    A pilot group, or a secondary domain where federation is domain-wide. A handful of accounts means the blast radius is a handful of accounts.

  4. Watch it answer real people

    Sign in from an approved machine, then an unapproved one, then one a version behind. Every attempt writes a record naming the condition that decided it, and that record is the artefact worth keeping.

Quick scan

The whole thing, on one screen

Decision point
At the identity provider, before the application opens
Decision type
Binary. No scoring, no weighting
Device established by
Binding, management enrolment, or agent posture
Applications
Any SAML application, plus private apps over the gateway
Posture checks
25 check types expressed as 144 named rules
Without an agent
Serial number, address range and location, at the gateway
Device approval
Administrator-held, per person, revocable in one place
Second factor
One condition among the set, scoped org-wide, per group or per user
On revocation
New sign-ins refused immediately; open sessions run to expiry
Record
One event per evaluation, naming the deciding condition, SIEM-exportable
Deployment
Gateway on a virtual machine inside the network
Asked on every evaluation

The questions that actually come up

Does this replace our device-management platform?

No, and it is not trying to. The management platform enrols devices and holds the compliance policies. Conditional access consumes that state and decides on top of it. If you have no management platform, binding and posture cover the same ground by different means.

We share desktops across shifts. Does that work?

Yes, and it is a common reason people arrive here. The machine is approved once, and each person who uses it is bound to it separately. One person holds a session at a time, so the desk is shared and the sessions stay apart.

Will people have to authenticate twice?

On machines where the agent carries the condition, the agent authenticates and so does the browser. That is deliberate on shared hardware, where a single sign-in would let whoever sits down next inherit the session. On personal and single-user machines it can be reduced to one.

Does the agent slow the machine down?

It is small, under ten megabytes, and it is reading state rather than inspecting traffic. It is not an endpoint protection product and does not behave like one.

Somebody leaves. What actually happens?

Suspend them in the console and every new sign-in is refused at once, across every device they hold. A browser session already open runs until it expires, which is a property of federated sign-in rather than a setting. Where the session has to end immediately, revoke the device as well.

Can we block phones by default and allow them on request?

Yes. New devices can be held pending approval as a rule, so a laptop can be approved and a phone left waiting until somebody asks for it. It is the same approval queue either way.

Which password policy applies if our suite has its own?

Whichever one the sign-in passes through. Once the domain is federated, the policy configured here is the one evaluated, so the two should be set to agree rather than left to differ quietly.

What about devices with nothing installed at all?

Three conditions are still enforceable at the gateway: a known serial number, a trusted address range and a permitted location. Fewer checks are available this way than with the agent, which is the honest trade-off for installing nothing. Device posture specifically needs the agent.

What exactly counts as compliant?

Whatever you define, from 25 check types expressed as 144 named rules: protection present and current, firewall state, disk encryption, operating system and patch level, domain membership, management enrolment, required and forbidden processes, and file and registry presence. A rule is a check, a threshold, and the group it applies to.

What does an auditor get?

One event per evaluation, against the connection, naming the condition that decided it rather than a bare allow or deny, exportable to a SIEM.

Bring the awkward population

Thirty minutes, against your own estate

Bring the people you cannot put a managed laptop in front of, and we will build the rule that lets them work anyway while you watch.

InstaSafe TechnologiesConditional access for managed, bound and unmanaged devices