Two different problems called identity
This course is about the identity of your own staff, contractors and systems. It is not about verifying customers, which is a different problem with different controls and a different body of guidance, and conflating the two is the first mistake.
Why they are different. Customer identity verification asks a question once, at the start of a relationship: is this person who they claim to be? Workforce identity asks a question continuously, for years: should this account still be able to do this thing today? The first is a proofing problem. The second is a lifecycle and authorisation problem, and almost all of its failures are failures of maintenance rather than of verification.
The vocabulary that keeps them apart. The current digital identity guidelines separate the question into levels that are selected independently of one another, which is itself the useful insight. Identity Assurance Level governs identity proofing — how sure you are who someone is. Authentication Assurance Level governs authentication — how sure you are that the person returning is the same one. Federation Assurance Level governs the federation case, where a relying party trusts an identity provider over a protocol. An organisation is "required to select an assurance level for each of the following functions" rather than buying one number for all three.
That separation matters practically. Your staff were proofed once, on their first day, often by a manager who recognised them. Their authentication happens forty times a week for the next six years. Spending on proofing and neglecting authentication is the standard shape of a workforce identity programme, and it is the wrong way round.
What the register has to answer. An identity programme is, in the end, a maintained answer to one question for every account in the organisation: what can this account do, who decided that, and when was it last checked? Three parts, and most programmes can produce only the first. If you cannot name the decision-maker, there is no accountability; if you cannot give the date, the entitlement is whatever it accumulated to.
Where the risk actually concentrates. Not in the ordinary user. In four places this course returns to repeatedly:
- Accounts that outlived their reason — the leaver who was disabled in the directory and not in the third-party system, the contractor whose engagement ended in March.
- Accounts with standing privilege — administrators who hold their rights permanently rather than when they need them.
- Non-human accounts — service accounts, API credentials and workload identities, which usually have no owner, no expiry and no review.
- The processes that change identity — password resets, authenticator enrolment, access requests. These are the controls attackers now target directly, and lesson five is a documented case of exactly that.
One scoping caution. Zero trust is the dominant frame for this subject, and it is a useful one, but a maturity model is not a programme. The point of the architecture is stated in the published guidance as protecting resources rather than network segments, with no implicit trust from location. Everything in this course either serves that or does not; vocabulary that serves neither is decoration.
Joiner, mover, leaver
Most organisations have a joiner process. Fewer have a mover process. Almost none can evidence a leaver process that covers everything the person could reach.
Why the mover case is the one that hurts. A joiner starts with nothing and is granted what the role needs — the default is safe. A leaver is a single event somebody is motivated to complete. A mover is neither: they gain the access of the new role and keep the access of the old one, because nobody is accountable for subtraction. Repeat that over a ten-year career and you have produced an account that can do almost anything, held by someone whose current job needs very little of it. Internally this is called accumulation; to an attacker it is simply a high-value target that nobody is watching.
The subtraction test. For any long-serving employee in a sensitive function, ask: what did they lose access to the last time they changed role? If the answer is "nothing" or "I would have to check", the mover process does not exist regardless of what the policy document says. This is a two-minute diagnostic and it is more informative than most audits.
Why leaver processes fail in a specific, predictable way. They are built around the directory, and the directory is not the whole estate. Disabling the central account closes the front door and leaves open:
- applications with their own local accounts, not federated to the directory;
- credentials that are not accounts at all — API keys, personal access tokens, SSH keys, signing keys the person generated;
- third-party and outsourced systems, where the person was provisioned by the supplier and your deprovisioning instruction is an email;
- shared credentials the person knew, which do not stop working when a person leaves;
- and enrolled authenticators that remain valid, including a device still in their possession.
The third-party point deserves its own emphasis, because the Cyber Safety Review Board's review of the Lapsus$ attacks found that organisations "did not always consider third parties and BPOs in their risk" programmes, while attackers specifically "targeted the organization's third-party servicers and business process outsourcers". If a business process outsourcer's staff hold accounts in your systems, their joiner-mover-leaver process is yours, and you are relying on it without operating it.
What good looks like, concretely. Joining, moving and leaving are triggered by the authoritative record in the human resources system rather than by a ticket somebody remembers to raise. Every application is either federated to the central directory or on an explicit list of exceptions with a named owner. Non-human credentials have owners and expiry dates, which is lesson ten. Contractor accounts carry an end date set at creation, so the default is expiry rather than permanence. And the leaver check is run against an inventory of systems rather than against the directory, because the directory will always say the job is done.
The number worth producing. How long, in days, between a leaver's last working day and the removal of their last piece of access — measured at the ninetieth percentile, not the average, because the average is dominated by the easy cases and the risk lives in the tail.
What "we have MFA" conceals
Multi-factor authentication is not one control. It is a family of controls with a range of resistance so wide that the single phrase covers both the strongest practical defence available and something attackers defeat routinely.
The published ranking. CISA's fact sheet on implementing phishing-resistant MFA sets out the forms from strongest to weakest, and the useful part is not the ranking but the specific threat each form does and does not resist:
- FIDO/WebAuthn authentication and public key infrastructure. Classed as phishing-resistant, described as "the gold standard for MFA". Push bombing, Signalling System 7 exploitation and SIM swap "are not applicable" to it.
- App-based one-time passcodes and push with number matching. "Vulnerable to phishing attacks", resistant to push bombing. SS7 and SIM swap are not applicable. The fact sheet names app or token-based one-time passcodes and push with number matching as "the best options for small- and medium-size business that cannot immediately implement phishing-resistant MFA".
- Push notification without number matching. "Vulnerable to push bombing attacks as well as user error."
- SMS or voice. "Vulnerable to phishing, SS7, and SIM swap attacks," and the guidance says this form "should only be used as a last resort MFA option."
Read what that list implies. The authenticator most organisations deployed — a six-digit code, from an app or a text message — sits in the two categories explicitly described as vulnerable to phishing. It defeats password spraying and credential stuffing, which is a real gain and the reason it was worth deploying. It does not defeat an attacker who stands up a convincing login page, because the user will type the code into it as readily as the password.
Why the difference is categorical rather than incremental. A one-time passcode is a secret the user can hand to anybody, including a website that is not yours. That is not a weakness in the implementation; it is what a transferable secret is. Phishing-resistant methods are not stronger secrets — they are a different mechanism, in which nothing transferable is produced at all. Lesson four is how that works, because understanding it changes how you evaluate every product claim in this space.
The honest position on migration. Phishing-resistant MFA is the goal and the guidance is unambiguous that it should be a priority. It is also a programme, not a setting: it involves hardware or platform authenticators, enrolment, recovery paths, and applications that may not support it. In the interim, the ranking tells you what to do in order — get off SMS and voice, turn on number matching where push is in use, and move the accounts that matter most first. Which accounts matter most is lesson eight.
The measurement. Not "do we have MFA", which is a yes. The number is: what proportion of our accounts, weighted by what they can reach, authenticate with a method the published guidance classes as phishing-resistant? For most organisations reading this honestly for the first time, the answer is a single-digit percentage concentrated in the security team.
Why origin binding is the whole trick
This lesson is the mechanism behind the word "phishing-resistant". It is worth understanding precisely, because once you have it you can evaluate any authentication claim in a procurement conversation without taking the vendor's word for anything.
The problem being solved. Every shared-secret authenticator — a password, a one-time code, a recovery phrase — can be given to the wrong recipient by a user who believes they are talking to you. No amount of user training removes this, because the user's belief is the thing being attacked, and a convincing replica of your login page is cheap.
The solution, in one sentence. Make the credential refuse to work anywhere but the site it was created for, and enforce that refusal in software the user cannot talk out of it.
How the specification does it. The Web Authentication specification, a W3C Recommendation at Level 3, scopes each credential to a relying party, and the key clause is where the enforcement lives: "This scoping is enforced jointly by conforming User Agents and authenticators", with the result that "the public key credential can only be accessed by origins belonging to that Relying Party." The browser and the authenticator check which site is asking. A replica site has a different origin, so the credential is not offered — not because the user was careful, but because the user was never consulted.
Why this is a structural property and not a feature. There is no secret to disclose. The authenticator holds a private key and produces a signature for a specific origin; nothing the user can read, type, screenshot or forward is sufficient to authenticate elsewhere. Push bombing has nothing to approve. SIM swapping has no message to intercept. That is why the published threat table says those attacks "are not applicable" rather than "are mitigated".
Where the standards draw the lines. The current authentication guidance ties this to assurance levels in a way worth knowing exactly:
- At the middle level, verifiers "SHALL offer at least one phishing-resistant authentication option", and the summary table marks phishing resistance as recommended and required to be available.
- At the highest level, authentication "requires a phishing-resistant authenticator" whose authentication key is non-exportable, and key exportability is prohibited.
- Consequently synced passkeys, which require the private key to be exportable in order to sync, "SHALL NOT be used" at the highest level — while being explicitly integrated into the guidance at lower levels.
The practical judgement that follows. Synced passkeys and hardware authenticators are both phishing-resistant and both defeat the attacks in lesson three. They differ on whether the key can leave the device, which matters when the threat you are defending against includes the compromise of the account the keys sync to. For the general workforce, syncable credentials remove the recovery and cost problems that stop hardware rollouts from finishing. For the small set of accounts that administer everything, the non-exportable version is the one the guidance reserves for the highest assurance, and that distinction is the right place to spend hardware budget.
What to ask a vendor. Not "do you support MFA" or "do you support passkeys". Ask: is the credential bound to our origin, is that binding enforced by the browser and the authenticator, and can the authentication key be exported? Three questions, and the answers are facts rather than positioning.
The help desk is an attack surface
In 2023 a statutory review board examined a set of intrusions in which the attackers were, by the standards of state-sponsored espionage, unsophisticated — and defeated the identity controls of large, well-resourced companies anyway. The findings are specific and they are about process, not technology.
How the multi-factor authentication was defeated. The Board recorded several distinct methods, and none of them involved breaking cryptography:
- Fatigue. Threat actors used "MFA fatigue, spamming employees with MFA prompts with the goal of overwhelming them with access approval requests until they said yes." The Board noted the prompts sometimes arrived "late at night, or during inconvenient times, possibly to increase the likelihood of the employee accepting."
- Impersonating the help desk. "In a few cases, the threat actor impersonated help desk personnel over direct chat messages and encouraged employees to approve the MFA prompts."
- Taking over the phone. Attackers used fraudulent SIM swaps to intercept passcodes and push notifications sent over SMS, which the Board described as "effectively defeating this widely used MFA control". It also noted that "a lucrative SIM swap criminal market further enabled this pay-for-access to a target's mobile phone services."
- Buying the credentials. Initial access brokers and the infostealer ecosystem — "whereby anyone can buy valid login credentials for a target" — were "highly effective means of initial entry".
The reverse direction, which is worse. A separate joint advisory on a related group, issued by seven national agencies and last revised in July 2025, documents the attack running the other way: the actors "posed as company IT and/or helpdesk staff using phone calls or SMS messages to obtain credentials from employees", and also "posed as employees to convince IT and/or helpdesk staff to provide sensitive information, reset the employee's password", including "spearphising calls to convince IT help desk personnel to reset passwords and/or transfer MFA tokens".
Read that last clause carefully. The attacker does not need to defeat your authentication if they can persuade your service desk to enrol a new authenticator for them. The reset process is the authentication control, and in most organisations it is operated by the lowest-paid, highest-turnover, most throughput-measured team in the business, against a verification script written for the problem of a forgetful employee rather than a hostile caller.
What the affected companies actually changed. The Board recorded the steps they took, and this is the most directly useful list in the report because it is what organisations did after learning the hard way: implementing phishing-resistant hardware tokens; eliminating MFA push alerts; verifying employee identities through video verification to reduce social engineering when fulfilling MFA reset requests; and transitioning to FIDO-supported applications and passwordless authentication.
The Board's recommendation, in its own terms. Organisations should "urgently implement improved access controls and authentication methods and transition away from voice and SMS-based MFA", and "should adopt easy-to-use, secure-by-default, passwordless solutions such as Fast IDentity Online (FIDO)2-compliant, phishing-resistant MFA methods". The review's framing sentence is worth keeping: "We need better technologies that move us towards a passwordless world, negating the effects of credential theft."
The action this lesson asks for. Read your own password reset and authenticator enrolment procedure as an attacker would. What does the agent verify, and could a well-prepared stranger supply it? Is there a step that cannot be socially engineered — a video check, a manager confirmation out of band, a wait? And is the agent permitted to decline a plausible, urgent, senior-sounding caller without being penalised for it? That last question is cultural and it decides the outcome.
Passwords, and what to stop doing
Password policy is the area where published guidance has moved furthest from common practice, and where the common practice is not merely dated but actively harmful. The current guidance prohibits several things most organisations still enforce.
What the guidance now says, in its own normative language. The authentication volume of the digital identity guidelines, final since July 2025, states plainly that "Passwords are not phishing-resistant." For password rules specifically: a single-factor password "SHALL ... be a minimum of 15 characters"; verifiers "SHALL NOT require subscribers to change passwords periodically"; and verifiers "SHALL NOT impose other composition rules" — the upper case, digit and symbol requirements that nearly every enterprise policy still contains.
Why forced rotation was abandoned. Not because it never helped, but because of what users do in response. A ninety-day expiry produces predictable transformations of one remembered password, written down more often and reused across systems, and it does nothing against the two threats that actually matter: a password phished today is used within minutes, and a password sold by an initial access broker is used while it is fresh. Rotation defends against slow offline cracking of a stolen hash, which is a narrow case, at the cost of degrading every password in the estate.
Why composition rules were abandoned. They reduce the search space rather than expanding it. When a policy requires one capital, one digit and one symbol, users satisfy it in the same few ways, and an attacker's wordlist encodes exactly those ways. Length does genuine work; mandated patterns mostly produce the illusion of it.
What to do instead, as a sequence. Raise the minimum length and allow long passphrases, including spaces. Remove periodic expiry, and replace it with rotation on evidence of compromise — which requires you to actually check against known-breached credentials. Keep the password out of the authentication decision wherever a phishing-resistant method is available, and recognise that the end state is not a better password but no password, which is what the review in lesson five recommended.
Session lifetime, which is the other half. Authentication is not a single moment; it is a moment plus a period during which the result is honoured. The guidance puts numbers on that period, and they are shorter than most estates use: at the middle assurance level, reauthentication should be no more than 24 hours overall and no more than 1 hour of inactivity; at the highest level, 12 hours and 15 minutes respectively. A thirty-day session on an administrative console undoes a strong authenticator, because the attacker who steals the session does not need to authenticate at all.
The uncomfortable organisational point. Removing forced password expiry is usually blocked not by security but by an audit checklist or a certification questionnaire that still asks for it. That conflict is resolvable — the current national guidance is the stronger authority and says so normatively — but it has to be argued rather than assumed, with the clause in hand. Write down which of your password rules exist because they reduce risk and which exist because a form asked for them; the second list is the one to take to your auditor.
Least privilege that survives the business
Least privilege is universally endorsed and rarely implemented, and the reason is not laziness. It is that the two standard ways of implementing it both fail at scale, in opposite directions.
Why role-based access control collapses. Group permissions into roles, assign roles to people: clean in theory. In practice every exception becomes a new role, because the alternative is granting someone more than they need. Organisations arrive at thousands of roles, many held by a single person, at which point the role model is a permission list with extra steps and nobody can say what any role means. This is role explosion, and it is the predictable consequence of modelling a changing organisation with a static structure.
Why the alternative is harder than it looks. The published definition of attribute-based access control is a methodology in which "authorization to perform a set of operations is determined by evaluating attributes associated with the subject, object, requested operations, and, in some cases, environment conditions" against policy. That expresses reality far better than a role does — this team, this data classification, this time, this device. The cost is in the phrase "evaluating attributes": it requires attributes that are accurate, maintained and available at the moment of decision. Most organisations do not have a reliable record of who reports to whom, so a policy that depends on it fails quietly and in the permissive direction.
The approach that works in practice. Not a choice between the two, but a division by stakes. Use coarse roles for the large population where the consequences of slight over-provisioning are tolerable, and reserve finely-specified, attribute-driven decisions for the systems where a wrong answer is serious. Trying to model the whole estate at the precision the crown jewels need is how these programmes stall for three years and deliver a spreadsheet.
The discipline that matters more than the model. Whatever the mechanism, three properties decide whether least privilege is real:
- Every entitlement has an owner who is accountable for whether it should still exist, and who is not the person holding it.
- Every entitlement has a reason recorded at the time it was granted — not a ticket number, a reason. Six months later the ticket says "approved" and nothing else, which is why reviews degenerate into guessing.
- Every entitlement has a default end. Permanent is a decision, and it should be an explicit one rather than the result of nobody specifying anything.
Where to start if the estate is a mess. Not with a complete model. Begin by finding the entitlements that are both widely held and highly consequential — the group that grants read access to everything, the role that can approve payments, the administrative group somebody added the whole department to in 2019. There are usually fewer than a dozen of these and they account for most of the real exposure. Fix those, with owners and dates, before attempting anything structural.
The measurement. Count the identities that could, today, read your most sensitive data store or make your most consequential change. Not the ones who should: the ones who could, including through nested group membership and through non-human accounts. In most organisations producing that number for the first time is itself the finding, and it is usually an order of magnitude larger than the owner expected.
Privileged access
Administrative access is where identity failures become incidents rather than findings, and the single most useful change is not a tool. It is removing the word "standing".
What standing privilege means and why it is the problem. An administrator who holds administrative rights continuously is an administrator whose credential is worth stealing continuously. The rights are present during the ninety-nine per cent of their week when they are reading email rather than administering anything, which is exactly when they are most likely to be phished. Nothing about their competence changes this; the exposure is a property of the arrangement.
The alternative, stated plainly. Privilege is requested, justified, time-boxed and recorded, so the account holds no administrative rights by default and acquires them for a defined window. This changes the economics for an attacker: a stolen credential yields an ordinary user, and the escalation step produces a record and a request that somebody can question.
The objections, and which of them are real. "It slows us down in an incident" — legitimate, and the answer is a break-glass path, which is the next paragraph. "Our administrators will not accept it" — usually a sequencing problem rather than a refusal; start with the least contested platform. "We cannot afford the tooling" — partly true, though the first eighty per cent is a decision about defaults rather than a purchase. The objection that is not real is "our administrators are trusted", because the control is not about their intentions.
Break-glass, done properly. There must be a path to privileged access that works when the normal path does not, or the first serious outage will produce a permanent exception. A sound break-glass account is one that is deliberately few in number, held with credentials stored outside the systems it recovers, excluded from ordinary federation so that an identity provider outage does not block it, monitored so that every single use generates an alert to someone other than the user, and tested on a schedule — because an untested emergency credential is usually an expired one.
The separation that gets forgotten. Administrative work should not be done from the same session, device or browser profile as general-purpose work. The compromise path in a large proportion of real intrusions is a document or a link opened in the same place the administrative console is signed in. This is a configuration decision available to every organisation at no cost, and it is skipped more often than any control in this course.
Who counts as privileged. Wider than the directory's administrator group. It includes whoever can change authentication policy, whoever can enrol or reset authenticators for others — lesson five's point, and the reason the service desk belongs in this category — whoever administers the code pipeline that deploys to production, and whoever holds the credentials of the non-human accounts in lesson ten. A privileged access programme scoped to the people whose job title says administrator has omitted most of its subject.
The measurement. How many identities hold administrative rights right now, as opposed to being able to request them? That number should be small, known, and falling. If it cannot be produced on demand, it is larger than anyone thinks.
Reviews that are not theatre
Access recertification is the most widely performed and least effective control in this field. Nearly every regulated organisation does it; very little access is ever removed by it.
Why it fails. A manager receives a list of two hundred entitlements with names like SAP_FI_GL_POST_02 and a deadline. They have no way to know what the entitlement grants, no record of why it was given, and no consequence for approving. Approving everything takes four minutes and challenging one item takes an afternoon of emails. The rational response is to approve, and that is what the data shows.
The signal that tells you it is theatre. The revocation rate. If a review cycle removes almost nothing, the review is not finding anything — and the correct inference is not that the estate is clean. It is the same reasoning as a human reviewer who never overrides an automated decision: a reviewer who never says no is not exercising judgement, and the control exists on paper only.
What changes the outcome. Four things, in order of effect:
- Describe the entitlement in business language. "Can post journal entries to the general ledger" rather than the technical identifier. This single change does more than any other, because it converts an unanswerable question into an answerable one.
- Show the reason and the age. Granted when, by whom, for what stated reason, last used when. Last-used data is decisive: an entitlement unused in a year is a safe removal and the reviewer can see it.
- Review the exposure, not the population. Reviewing every entitlement annually is a volume exercise that guarantees rubber-stamping. Review the consequential entitlements quarterly and in detail, and let the rest be driven by the lifecycle in lesson two.
- Default to removal. If a reviewer does not respond, the entitlement lapses rather than persisting. This inverts the incentive, and it is the one change organisations resist most, because it will break something the first time it runs. It will also be the first time anyone finds out who actually needs the access.
The honest cost. Default-to-removal creates operational breakage and a wave of re-request work. That is the price of discovering the truth about an estate, and it is cheaper paid deliberately in a review cycle than involuntarily during an incident. Run it on one application first, measure how much was re-requested, and you will have both the evidence and the objections for the wider rollout.
What to produce for whoever asks. Not the completion percentage, which only measures whether managers clicked. The pair of numbers that mean something: what proportion of reviewed entitlements were revoked, and what proportion of entitlements in the estate were in scope at all.
Machines have identities too
Non-human identities now outnumber human ones in most estates, and they are governed far less well: no joiner-mover-leaver, no review, frequently no owner, and credentials that never expire.
Why the old method stopped working. Service-to-service authorisation used to be expressed as network facts — this subnet may call that address. The zero trust guidance for cloud-native applications sets out why that fails now: network parameters are "insufficient", and "the presence of proxies, network address translations, dynamic infrastructures (e.g., migration of applications between VMs), and load balancers make it impossible for the called application to know the IP address of the calling application in order to make authentication or authorization decisions."
What replaces it. Policies expressed in identities rather than addresses. The guidance contrasts the two directly: an identity-tier policy reads like "order processing front-end service can communicate with inventory back-end service", where a network-tier policy reads "permit calls from 192.168.10.0/24 subnet to 10.0.0.31". The first survives a redeployment; the second has to be maintained in step with infrastructure that changes for reasons unrelated to security.
What it requires. The guidance names three prerequisites, and they are a useful checklist for anything claiming to do this:
- A unique identity for each application or service, "just like how each user carries a unique identity".
- Trust in that identity — it "should not be subject to spoofing and should be continuously verifiable". The worked example is a SPIFFE ID, a URI-encoded workload identifier carried in a cryptographically verifiable document, most commonly an X.509 certificate.
- Discovery of application resources, so you know every dependency consumed over the network, "enabled through an authenticated service registry".
The method for organisations that do not know their own call graph, which is most of them. The guidance is refreshingly practical: observability data such as service mesh metrics can build a view of "requests made today", from which fine-grained policies are written. It calls this an "observe-and-lock-down methodology", and notes that doing it builds the organisational processes needed to maintain those policies over time. Start from what is happening, not from a design document.
A caution on tooling claims. Container network policy implemented at the network layer, via mechanisms like eBPF, provides identity-style policy and the guidance calls it "a strong upgrade from traditional microsegmentation" with finer-grained policies that are easier to manage. But it notes these approaches "typically lack the ability to apply per-request policies in the context of the application". Better than subnets, not the same as identity-based segmentation — and a vendor conflating the two is worth questioning.
The mundane half that causes most incidents. Secrets. An API key in a repository, a token in an environment variable in a wiki page, a service account whose password is in a runbook. Every non-human credential needs a named human owner, an expiry, a rotation method that has been executed at least once, and a place to live that is not a source file. Lesson two's leaver problem applies here too: when the owner leaves, the credential does not.
Sessions, tokens and federation
Authentication produces a token, and the token is what the attacker actually wants. A programme that hardens the login and ignores what follows has protected the door and left the key in it.
Why token theft is now the common path. If a phishing-resistant authenticator makes credential capture useless, the attacker's next target is the artefact issued after a successful authentication — a session cookie or a bearer token. This is not hypothetical: the cloud matrix of the standard adversary technique catalogue lists stealing application access tokens, and stealing or forging authentication certificates, as named techniques precisely because they bypass the authentication the organisation invested in.
What makes a stolen token valuable. Lifetime and scope. A token that is valid for thirty days, accepted from any device and any network, and scoped to everything the user can do, is a complete bypass of every control at the front door. The same token with a one-hour life, bound to the device that requested it, and scoped to one application, is a much smaller prize. These are configuration decisions, usually left at the vendor's default because shortening them generates help desk calls.
The numbers the guidance gives. Reauthentication at the middle assurance level should be no more than 24 hours overall and 1 hour of inactivity; at the highest level, 12 hours and 15 minutes. Compare those against the session lifetimes in your own administrative tools. The gap is typically large, and it is the cheapest risk reduction available in this course.
Federation, and the assurance level nobody selects. When a relying party trusts an identity provider over a protocol, the strength of that link is its own question — which is why the guidance defines a Federation Assurance Level separately from proofing and authentication. In practice almost nobody chooses one deliberately. The consequences of not choosing: an application that accepts assertions without checking the audience they were issued for, trust relationships nobody can enumerate, and a federation configuration that no longer matches the applications actually using it.
The concentration this creates. Federating everything to one identity provider is correct — it is how lesson two's leaver process becomes enforceable — and it also means that provider's compromise is the compromise of everything behind it. The documented Exchange Online intrusion is the clearest illustration: a signing key that an adversary could use to mint tokens accepted across a service, which the Board described as giving "full access to essentially any Exchange Online account anywhere in the world". Centralise, and then treat the centre as the most consequential system you operate.
What to check this week. List the session and token lifetimes for your administrative consoles. Find out whether tokens are bound to a device or merely to a user. Enumerate your federation trusts and identify who owns each one. And establish whether you can revoke a session immediately — not disable the account, revoke the live session — because during an incident that is the control you will reach for and the one most often missing.
What to measure
Identity programmes run for years and report activity rather than outcome. This lesson is what to put in front of a board instead, and how to think about maturity without mistaking a model for a plan.
The maturity frame, used correctly. The published zero trust maturity model is built on five pillars — Identity, Devices, Networks, Applications and Workloads, and Data — with three cross-cutting capabilities: Visibility and Analytics, Automation and Orchestration, and Governance. Progress within each is described across four stages: traditional, initial, advanced and optimal. That structure is genuinely useful for locating yourself and for showing that identity is one pillar of five rather than the whole programme.
Its limit, stated plainly. A maturity model describes a destination; it does not tell you which of your own systems is the most dangerous today. An organisation that moves every pillar from traditional to initial has spread a thin layer of improvement across everything and may not have fixed a single thing that would have stopped the attacks in lesson five. Use the model to report position and the measurements below to decide work.
The numbers worth reporting. Each of these has appeared in a lesson above, and each is a fact rather than an opinion:
- Proportion of accounts, weighted by what they can reach, using a method the guidance classes as phishing-resistant.
- Number of identities holding standing administrative rights, as against being able to request them.
- Days from a leaver's last working day to removal of their last access, at the ninetieth percentile.
- Revocation rate in access reviews, alongside the proportion of the estate in scope.
- Count of identities that could reach the most sensitive data store, including through nested groups and non-human accounts.
- Number of non-human credentials with no named owner or no expiry.
- Session and token lifetimes for administrative consoles, against the published reauthentication figures.
Why these and not the usual ones. Because each one moves only if something real changed, and each one is uncomfortable the first time it is produced. Completion percentages, training rates and tool deployment counts all rise without any attacker's job becoming harder.
The three things this course asks you to do first, if the list is too long to start everywhere. Read your own authenticator reset procedure as a hostile caller would, because that is the control the documented cases actually attacked. Remove standing administrative rights from one platform, because it changes the value of every stolen credential. And produce the leaver tail — the ninetieth percentile, measured against an inventory of systems rather than the directory — because it is the number most likely to be far worse than anyone in the organisation currently believes.
A closing observation. Nothing in this course required a technology that does not exist. The intrusions reviewed by a statutory board were carried out against well-resourced companies by actors the Board itself did not describe as sophisticated, and the remediations those companies adopted afterwards — hardware tokens, no push alerts, video verification on reset requests, passwordless authentication — were available beforehand. The gap was not capability. It was that nobody had read the reset procedure as an attacker, and nobody had asked how many people held standing privilege.
Fifty questions
Answer all fifty questions, then check your answers. You need 45 out of 50 to pass and receive a certificate. If you score less, you will see which answers were right and wrong, and then go through the course again before you retake the check. Your answers, progress and times are kept only in this browser.
Your answers
Your certificate of completion
Enter your name as you want it to appear, then save the certificate as a PDF. In the print window, choose Save as PDF. A certificate is issued once per completion of the course.
Rate this course
How would you rate this course? Your rating is anonymous — just a number, with no name and no account.
Bring this to your team
Your colleagues can take this course too — it is free, needs no sign-up, and nothing they enter leaves their browser. If your institution wants the same ground covered as structured training for a team, with an instructor and your own cases, we run courses in a classroom or online for groups of ten or more.
Talk to us about training your team → See the paid courses →
Saolix does not record who takes this course, so it cannot verify these certificates. The certificate confirms completion of a free self-paced course and is not an accredited qualification.
Standards and frameworks covered
The published frameworks this course teaches toward. Saolix is not affiliated with, endorsed by or accredited by any of these bodies; naming a standard is a citation, not a claim about its author.
- NIST — NIST SP 800-63-4: Digital Identity Guidelines, NIST SP 800-63B: Authentication and Authenticator Management, NIST SP 800-207: Zero Trust Architecture, NIST SP 800-207A: A Zero Trust Architecture Model for Access Control in Cloud-Native Applications in Multi-Cloud Environments, NIST SP 800-162: Guide to Attribute Based Access Control (ABAC) Definition and Considerations, NIST SP 800-210: General Access Control Guidance for Cloud Systems, NIST SP 800-57 Part 1 Revision 5: Recommendation for Key Management, NIST SP 800-61 Revision 3: Incident Response Recommendations and Considerations for Cybersecurity Risk Management
- W3C — Web Authentication: An API for accessing Public Key Credentials, Level 3
- MITRE ATT&CK — ATT&CK cloud matrix
- ENISA — ENISA Threat Landscape 2025
Sources
The official documents this course relies on. Laws and guidance change, so check the current version.
- NIST SP 800-63-4: Digital Identity Guidelines · US National Institute of Standards and Technology
- NIST SP 800-63B: Authentication and Authenticator Management · US National Institute of Standards and Technology
- Review of the Attacks Associated with Lapsus$ and Related Threat Groups · Cyber Safety Review Board, US Department of Homeland Security
- Implementing Phishing-Resistant MFA · US Cybersecurity and Infrastructure Security Agency
- Scattered Spider (joint advisory AA23-320A) · FBI, CISA, RCMP, Australian Signals Directorate, Australian Federal Police, Canadian Centre for Cyber Security and NCSC-UK
- Web Authentication: An API for accessing Public Key Credentials, Level 3 · World Wide Web Consortium
- Zero Trust Maturity Model, version 2.0 · US Cybersecurity and Infrastructure Security Agency
- NIST SP 800-207: Zero Trust Architecture · US National Institute of Standards and Technology
- NIST SP 800-207A: A Zero Trust Architecture Model for Access Control in Cloud-Native Applications in Multi-Cloud Environments · US National Institute of Standards and Technology
- NIST SP 800-162: Guide to Attribute Based Access Control (ABAC) Definition and Considerations · US National Institute of Standards and Technology
- NIST SP 800-210: General Access Control Guidance for Cloud Systems · US National Institute of Standards and Technology
- NIST SP 800-57 Part 1 Revision 5: Recommendation for Key Management · US National Institute of Standards and Technology
- NIST SP 800-61 Revision 3: Incident Response Recommendations and Considerations for Cybersecurity Risk Management · US National Institute of Standards and Technology
- Review of the Summer 2023 Microsoft Exchange Online Intrusion · Cyber Safety Review Board, US Department of Homeland Security
- ATT&CK cloud matrix · The MITRE Corporation
- Cloud security guidance · UK National Cyber Security Centre
- ENISA Threat Landscape 2025 · European Union Agency for Cybersecurity
- SS2/21: Outsourcing and third party risk management · Bank of England, Prudential Regulation Authority



