The line you cannot move
Every cloud contract divides the work of security between you and the provider. Most organisations can recite the diagram and almost none can say what it leaves them holding.
What the model is for. The shared responsibility model exists because you cannot audit a hyperscale data centre and the provider cannot know what your data means. So the provider takes the parts that are uniform across every customer — the physical estate, the hypervisor, the managed service's own code — and you take the parts that are specific to you: who may log in, what is exposed, what the data is, how long you keep it, and whether you would notice if it left.
Where the line actually sits. It moves with the service model, and it moves further than people expect. NIST's access control guidance for cloud systems makes the point precisely: guidance written for infrastructure as a service also applies to platform and software as a service, and guidance for infrastructure and platform also applies to software as a service. Responsibility accumulates downward. Buying a more managed service removes work from your side of the line; it never removes accountability from your side of the contract.
The three sentences that find the gap. For any system in your estate, answer these without looking anything up:
- If this service were misconfigured tomorrow, who would be told, and how?
- If an attacker authenticated as a legitimate user, which of our logs would show it?
- If the provider had an incident, what would we be able to establish on our own?
An estate where those answers are confident is a governed estate. An estate where they are "the provider handles that" has misread the diagram, because none of the three is on the provider's side of any shared responsibility model in commercial use.
Why this is a cyber course and not a procurement one. CISA's #StopRansomware guide, a document about defending against a specific criminal business model, still tells its readers to "review the shared responsibility model for cloud and ensure you understand what makes up customer responsibility when it comes to asset protection." That instruction appears in a ransomware playbook because the gap it describes is where incidents happen — not in the provider's failure to patch a hypervisor, but in the customer's assumption that someone else was watching.
What the rest of this course does. It works through the customer's side of the line in the order that incidents actually travel: identity first, because that is now the perimeter; then keys, evidence, configuration and data; then the workload; then what you do when something has already gone wrong and the logs belong to somebody else.
Identity is the control plane
In a data centre, an attacker who wanted your database had to reach it across a network you controlled. In a cloud, they need a token.
What changed. The network is no longer the thing that decides. NIST's zero trust architecture publication states the principle plainly: zero trust "assumes there is no implicit trust granted to assets or user accounts based solely on their physical or network location", and it "focuses on protecting resources (assets, services, workflows, network accounts, etc.), not network segments." That is not a product description. It is an observation that location stopped being evidence of anything the moment the resource moved to somebody else's network.
What replaces it. Authentication and authorisation become discrete decisions taken before each session, for both the subject and the device. In practice this means the identity provider — not the firewall — is the system whose compromise ends the argument. If an attacker can mint credentials your services accept, every network control you own is downstream of a decision already made in their favour.
What this looks like to an attacker. MITRE's ATT&CK cloud matrix covers four platforms: identity provider, SaaS, infrastructure as a service, and office suite. Read the technique list and the shape of cloud attacks is obvious, because the named techniques are not about breaking in from outside:
- Steal Application Access Token (T1528) and Steal or Forge Authentication Certificates (T1649) — take the thing that proves identity rather than the credential a human types.
- Modify Authentication Process, with sub-techniques for multi-factor authentication (T1556.006), hybrid identity (T1556.007) and conditional access policies (T1556.009) — change the rules that decide who is trusted.
- Cloud Instance Metadata API (T1552.005) — ask the platform itself for the credentials attached to a running workload.
- Modify Cloud Compute Infrastructure (T1578), whose sub-techniques include creating and reverting snapshots and instances — use the control plane's own legitimate operations.
Every one of those is an authenticated action. None of them trips a perimeter control, because by the time they occur the actor is inside the identity system's idea of normal.
What follows for a defender. Three consequences, in order of how often they are missed. First, the identity provider is a tier-zero system and should be governed like a domain controller, not like an application. Second, conditional access and MFA configuration are security controls whose change history matters as much as their current state — T1556.009 exists because policies get edited, not bypassed. Third, workload identity deserves the same scrutiny as human identity: a metadata service that will hand credentials to anything running on the instance means a single application vulnerability yields platform access.
The uncomfortable implication. If identity is the control plane, then the most consequential thing in your estate is a key or a token, and the next two lessons are about what happens when one of those is taken.
What a documented failure looks like
Security courses are full of hypotheticals. This lesson uses an incident that was investigated by a statutory board, reported in public, and is specific enough to argue with.
The facts. In May and June 2023 a threat actor compromised the Microsoft Exchange Online mailboxes of 22 organisations and over 500 individuals. The actor, Storm-0558, was assessed as affiliated with the People's Republic of China and pursuing espionage. It accessed the accounts using authentication tokens signed by a key Microsoft had created in 2016. The victims included the mailboxes of a serving US Commerce Secretary, the US Ambassador to China and a member of Congress. The United States Cyber Safety Review Board reviewed the incident and published its findings on 20 March 2024.
What made it severe. The Board's framing is worth quoting because it is the whole lesson in one sentence: signing keys "are the cryptographic equivalent of crown jewels for any cloud service provider", and "an adversary in possession of a valid signing key can grant itself permission to access any information or systems within that key's domain." In this case the key's reach, combined with a separate flaw in the authentication system, "permitted Storm-0558 to gain full access to essentially any Exchange Online account anywhere in the world."
The finding that should change how you buy cloud services. As of the report's publication, almost a year after the intrusion, "Microsoft does not know how or when Storm-0558 obtained the signing key." The Board concluded the intrusion "was preventable and should never have occurred", and that the provider's "security culture was inadequate and requires an overhaul". Among the seven grounds it gave, the second is the one defenders should sit with: the provider's "failure to detect the compromise of its cryptographic crown jewels on its own, relying instead on a customer to reach out to identify anomalies the customer had observed."
Read that again from your side of the line. The provider did not catch it. A customer did. The Board also recorded that it assessed security practices at other cloud providers and found controls that this one did not maintain — so this was not an industry-wide limit on what is possible, but a gap specific to one provider that its customers had no way to see.
What it is not. It is not an argument against cloud computing, and it is not an argument about one company. The Board, whose mandate covered "a broader review of issues relating to cloud-based identity and authentication infrastructure affecting applicable cloud service providers and their customers", issued recommendations to every cloud provider, not only to the one it investigated. Those recommendations cover identity controls informed by a rigorous threat model, a minimum standard for default audit logging, emerging digital identity standards, incident and vulnerability disclosure, and victim notification.
The honest conclusion to draw. You cannot verify a provider's internal controls, and the one organisation that could not explain how its most sensitive key was taken was a company with enormous security resources. Therefore the question for your own estate is not whether a provider will fail, but whether you would be able to tell, and what you would be able to establish afterwards. That question is answered by your logging, which is lesson five, and by your contracts, which is lesson twelve.
Keys, and who holds them
The previous lesson turned on a key created in 2016 that was still signing tokens accepted in 2023. That is not only an operational failure; it is a departure from published guidance that is worth understanding properly, because the same reasoning applies to every key in your own estate.
What a cryptoperiod is. NIST's key management recommendation defines it as "the time span during which a specific key is authorized for use by legitimate entities or the keys for a given system will remain in effect." Keys are not meant to be permanent. They are meant to expire.
Why it is bounded. The guidance gives explicit reasons, and they read as a direct description of what went wrong in the Exchange Online case. A suitably defined cryptoperiod:
- limits the amount of information available for cryptanalysis to reveal the key;
- limits the amount of exposure if a single key is compromised;
- limits the use of a particular algorithm to its estimated effective lifetime;
- limits the time available for attempts to penetrate the physical, procedural and logical access mechanisms that protect a key; and
- limits the period within which information may be compromised by inadvertent disclosure.
The second and fourth are the ones to hold onto. A key with no expiry gives an adversary unlimited time to attempt the theft, and unlimited reach once it succeeds.
The counter-argument, which the same document makes. Shorter is not automatically better. The guidance notes that "short cryptoperiods may be counter-productive, particularly where denial-of-service is the paramount concern and there is a significant potential for error in the re-keying or key-derivation process." Rotation is an operation that can fail, and a failed rotation is an outage. It also observes that keys protecting stored data generally get longer cryptoperiods than keys protecting communications, because re-encrypting everything at rest is expensive. This is a judgement, not a number, and anybody who gives you a single universal rotation interval has not read the source.
The questions to ask about your own keys. Not "do we rotate?" but:
- For each key that signs or encrypts something, what is its defined cryptoperiod, and who decided it?
- If it had to be rotated unexpectedly today, who would do it, has it been rehearsed, and what breaks?
- Which keys are held by us, which by the provider, and for the provider's keys, what do we actually know about their management?
The last question is the hard one. For most managed services you will not be told, and the Board's review demonstrates that the answer can be unsatisfactory even at the largest providers. Where the data genuinely warrants it, customer-managed keys move that control to your side of the line — at the cost of taking on the rotation, availability and recovery problem yourself, which is a real cost and not a formality. Choosing to leave keys with the provider is a legitimate decision. Making that choice without recording it as a decision is not.
Logging you can investigate with
This is the lesson that changes outcomes, and it is the one most often deferred until an incident proves it was needed.
How the Exchange Online intrusion was found. Not by the provider. The US State Department's security operations centre detected anomalies in access to its mail systems on 15 June 2023, and the next day saw multiple alerts from a custom rule it had written and named internally "Big Yellow Taxi". That rule analysed a log called MailItemsAccessed, which records access to mailboxes. State could build the rule because it had bought a licence tier that included enhanced logging. The Cyber Safety Review Board recorded the consequence plainly: that log "was not accessible without that 'premium' service."
Three facts from that sequence, each worth a control.
First, the log that caught a nation-state actor was behind a paywall. Few of the victims had bought it. The Board recommended that cloud providers "adopt a minimum standard for default audit logging in cloud services to enable the detection, prevention, and investigation of intrusions as a baseline and routine service offering without additional charge" — which tells you it was not a baseline at the time.
Second, buying the log would not have been enough. The Board heard that few organisations analysed the voluminous MailItemsAccessed log in detail, and noted such analysis "would be difficult for smaller organizations." State found the intrusion because its SOC had three years of experience with anomalous mailbox access and had used it to build rules that enumerated every application touching a mailbox and alerted on deviations from a baseline. The log was necessary; the engineering on top of it was what worked.
Third, and most expensive to learn late: logging is not retroactive. When victims without the premium licence were identified, the provider switched enhanced logging on for them — but could not give them historical logs. They could see from the moment logging was enabled and no earlier. The intrusion window stayed dark. One victim organisation's assessment of its own incident was impaired by an incomplete dataset it had no way to complete.
What to do with that. The question is not "do we have logging?" It is a short list you can answer this week:
- Which log would show an attacker authenticating as a legitimate user, and do we have it today, at the tier that retains it?
- What is the retention period, and is it longer than the time we would plausibly take to notice? Intrusions measured in months are routine; thirty days of logs are not an investigation.
- Who has written a detection rule against these logs, and when was it last tuned? A log nobody queries is storage, not security.
- If the provider enabled additional logging for us tomorrow, how far back would it reach? If the answer is "it would not", that is the gap.
The uncomfortable economics. Enhanced logging tiers cost money, the volume is large, and the benefit is invisible until the day it is the only thing you have. That is a budget argument, not a technical one, and it is best made before the incident with the Board's finding in hand: the only reason the world learned about this intrusion is that one customer had paid for the log and built something on top of it.
Configuration, and holding a baseline
Most cloud incidents do not involve a clever attack. They involve a setting.
Why defaults matter more than policies. The UK National Cyber Security Centre's cloud security guidance is goal-based rather than prescriptive, and one of its goals is worth adopting as a procurement test on its own: prefer a service that "makes configurable security-enhancing features opt-out, and not opt-in". A control that must be switched on will be missing somewhere in your estate. A control that must be switched off will be present almost everywhere.
The trap inside improving defaults. The same guidance records something that catches experienced teams: "Some cloud platforms have adopted a more secure by default approach recently, and only apply those good defaults to new accounts or tenants." Your oldest and most important tenant is the one least likely to have the new protections, because the improvement was never applied retroactively. The guidance's answer is to "periodically review your service configuration against the platform's good practice guides", on the explicit basis that newer mitigations "may not be retroactively applied".
That is a scheduled task, not a project. An estate audited once at migration and never again is configured as the platform was in the year you migrated.
What to measure against. You need an external baseline, because an internally written standard drifts toward whatever the estate already does. Two kinds are published and free to use:
- Benchmarks such as the Center for Internet Security's, which are consensus configuration recommendations covering more than twenty-five vendor product families and include foundations benchmarks for the major cloud platforms, each versioned so you can state which version you assessed against.
- Baselines with tooling, such as CISA's Secure Cloud Business Applications project, established in 2022 "to address cybersecurity and visibility gaps exposed by software-as-a-service (SaaS) cyber intrusions and compromises". It publishes secure configuration baselines and free assessment tools that check a tenant's configuration against them. The significant part is not the tool; it is that the baseline is written down by somebody other than you, so a gap is a finding rather than an opinion.
Making it hold. The NCSC's guidance also sets the goal that "all service configuration can be set and audited using infrastructure as code, or via an API". That is the difference between a baseline and a baseline that survives. If configuration is code, then drift is a diff, a change has an author and a reviewer, and the current state can be re-derived rather than remembered. If it is console clicks, the baseline is accurate on the day of the audit and unknown the day after.
The honest caveat. Infrastructure as code moves the risk rather than removing it: the templates become the thing worth attacking, and a permissive template applied consistently is a misconfiguration deployed at scale. CISA's ransomware guidance accordingly says that infrastructure-as-code "should be version controlled and changes to the templates should be audited", and that copies of template files should be kept offline so resources can be redeployed quickly. Treat the repository that defines your estate as part of the estate.
Data, encryption, and what it does not solve
Encryption is the control most often cited and least often examined. It is necessary, it is cheap, and it answers a narrower question than people assume.
What the question actually is. The NCSC's lightweight approach to cloud security names four areas where you depend on the provider to protect your data: data encryption, authentication and access control, security logging and incident management, and governance. Encryption is one of four, and it is listed alongside access control for a reason — it protects data from someone who obtains the storage, not from someone who obtains an identity with permission to read it.
The failure mode this explains. Almost every cloud data loss you have read about involved correctly encrypted data. Encryption at rest was on; the attacker authenticated as something entitled to the plaintext and the platform decrypted it for them, as designed. Encryption at rest defeats a stolen disk, a decommissioned drive and a provider employee with physical access. It does not defeat a stolen token, and the previous lessons established that a stolen token is the normal shape of a cloud attack.
So what does encryption buy you? Three things worth paying for, stated honestly:
- It removes a whole class of physical and residual-data risk you cannot otherwise manage in someone else's data centre.
- With customer-managed keys, it gives you a revocation lever: destroy or withhold the key and the data is unreadable even to the platform. That is a real capability in an exit or a breach, and it is the strongest argument for taking on key custody.
- It is frequently a legal or contractual requirement, and meeting it is not optional even where the threat model is thin.
What it does not buy is protection from your own authorisation model. That is the access control half of the pair.
The questions that find real data risk. Not "is it encrypted?" — the answer is almost always yes and tells you almost nothing. Ask instead:
- Where is it? Not the region on the contract: the actual list of stores. Object storage, the database, the warehouse, the search index, the message queue, the backups, the non-production copy somebody took for testing, the export in a spreadsheet in someone's drive. Data has copies, and the copies rarely inherit the controls.
- Who can read it? Enumerate the identities with the permission, including service identities and anything with a standing administrative role. The number is usually larger than the owner believes.
- Would we know if it were read? This is lesson five's question applied to data rather than to logins, and for most data stores the honest answer is no, because object-level read logging is a tier you have to buy.
- How long do we keep it, and does the retention actually execute? A deletion policy that no job enforces is a document, not a control, and it enlarges every future breach.
The point to take away. Encryption is table stakes and should be on everywhere, with keys whose cryptoperiod you have decided. The work that reduces data risk is the inventory, the permission review and the read logging — and none of those is a cryptographic problem.
Segmentation when there is no perimeter
Network controls did not stop mattering in the cloud. They stopped being the primary control, which is a different claim, and the distinction is where teams go wrong in both directions.
The direction most teams get wrong first. Having read that identity is the new perimeter, they stop designing the network. Then a single compromised workload can reach everything, because every subnet can route to every other one and egress is unrestricted. Identity being primary does not make the network free; it makes the network the control that limits what a successful identity compromise reaches.
The specific thing to defend. Lesson two listed Cloud Instance Metadata API (T1552.005) among the techniques in ATT&CK's cloud matrix. It deserves its own paragraph because it is the clearest example of why workload-level network control still matters. A workload in a cloud usually has an identity, and the platform exposes that identity's credentials to the workload over a local service. An application vulnerability that lets an attacker make requests from the host — a server-side request forgery, a vulnerable dependency, an exposed debug endpoint — can therefore become platform credentials. The network path from the application to the metadata service is the whole attack, and restricting it is a network control protecting an identity control.
Egress is the half nobody designs. Inbound rules get written because somebody has to make the service work. Outbound rules get left open because nothing breaks when they are. But exfiltration is outbound, command and control is outbound, and the retrieval of the attacker's tooling is outbound. An estate with careful ingress and unrestricted egress has instrumented the half of the conversation that matters least.
Inside the workload. NIST's container security guide names poorly separated inter-container network traffic and the mixing of workload sensitivity levels among its orchestrator risks, and unbounded network access from containers among its container risks. Those are three ways of describing one failure: a flat internal network inside a cluster, where the public-facing component and the component holding the data are one hop apart and nothing records the hop.
What good looks like, without naming a product.
- Workloads of different sensitivity are separated, and the separation is enforced somewhere the application cannot alter.
- Egress is default-deny with an allow list, for the segments where that is achievable, and is at minimum logged everywhere else.
- Access to the metadata service is restricted to what genuinely needs it, and the workload identity's permissions are scoped on the assumption that the metadata service will eventually be reached.
- Administrative access to the control plane does not traverse the same path as ordinary user traffic.
- Every rule is expressed as code, for the reasons in lesson six.
The judgement to apply. Zero trust publications are explicit that the model protects resources rather than network segments. That is a statement about where trust is decided, not permission to flatten the network. The useful reading is: never treat a network location as proof of identity, and still make sure that an attacker who proves identity falsely has somewhere small to stand.
The workload, the image and the registry
Everything so far has concerned the platform. This lesson concerns what you put on it, because a hardened account running a compromised image is a hardened account running a compromised image.
Why containers changed the problem. NIST's application container security guide sorts the risks into five layers — images, registries, orchestrators, containers and the host operating system — and the value of that structure is that it separates problems teams routinely conflate. Scanning an image does nothing about an over-privileged orchestrator. Hardening the host does nothing about a secret baked into a layer.
The image risks worth internalising. The guide names image vulnerabilities, configuration defects, embedded malware, embedded clear-text secrets and the use of untrusted images. Two of those five are not about vulnerabilities at all:
- Embedded clear-text secrets. A credential written into an image is in every copy of that image, in the registry, in every running container and in the layer history even if a later layer deletes it. Image scanning for known vulnerabilities will not find it. This is the most common way a long-lived credential escapes.
- Untrusted images. Most images are assembled from a base somebody else published. Your supply chain is as long as your base image's, and the question "who can publish to the registry we deploy from" is a more important access control decision than it is usually treated as.
The registry risks are access control problems. The guide names insecure connections to registries, stale images in registries, and insufficient authentication and authorisation restrictions. Stale images are the quiet one: the vulnerability was fixed in the image six months ago, and the thing actually running was pulled by digest or tag before the fix. An image inventory that lists what you have built is not an inventory of what you are running.
The orchestrator is a control plane of its own. Its named risks include unbounded administrative access, unauthorised access, poorly separated inter-container traffic, mixing of workload sensitivity levels, and orchestrator node trust. Note what that list is: the same identity-and-segmentation argument as lessons two and eight, at a different layer. A cluster administrator is a platform administrator for everything in the cluster, and the credential that deploys is frequently more powerful than the credential that administers the cloud account.
A word about the guide's date. It was published in September 2017, and the tooling has moved considerably since. The risk taxonomy has not, which is why it remains worth reading: the categories describe where things go wrong structurally, and a 2017 document that still maps cleanly onto a current estate is telling you the problems are architectural rather than technological.
The practical sequence. Decide who may publish to the registry you deploy from. Scan images, but separately scan for secrets, because the two scans look for different things. Pin what you deploy and know what is running rather than what was built. Treat the orchestrator's administrative access as tier zero. And accept that this layer is where the fastest-moving risk in your estate lives, because it is the layer your delivery teams change daily.
Responding when the evidence is not yours
An incident in your own data centre is an investigation you can run. An incident in a cloud is an investigation you run partly through someone else's support process, and the difference is best understood before you need it.
What the documented case showed. The Cyber Safety Review Board recorded that victims "found it difficult to investigate these intrusions after initial detection because Microsoft could not, or in some cases did not, provide victim organizations with holistic visibility into all necessary data." One victim asked for the logs the provider had used to reach its own conclusions, received only portions of its audit logs and a set of IP addresses to search against, and the Board found that "this incomplete dataset impacted" that organisation's ability to complete its assessment.
The three asymmetries that creates. You have less data than the provider; you get it later than the provider; and you may not get all of it. None of those is fixable during an incident. All three are addressable before one:
- Data. Covered in lesson five. The controllable variable is which logging tier you hold today and for how long it retains.
- Timing. Know the provider's notification commitment and the route by which it arrives, and know who in your organisation receives it out of hours. The Board thought this weak enough across the industry to recommend that providers "develop more effective victim notification and support mechanisms".
- Completeness. Ask, as a procurement question rather than an incident question, what forensic data the provider will release to a customer under investigation, in what format and on what timescale. A contract that is silent on this is a contract that resolves it at the provider's discretion.
Where the published guidance now points. NIST's incident response recommendations were revised in April 2025, replacing a document from 2012, and the revision is framed as a Community Profile of the Cybersecurity Framework 2.0 — that is, incident response treated as part of continuous risk management rather than as a standalone lifecycle run by a separate team. For a cloud estate that reframing is not academic. The things that determine whether you can respond — logging tiers, contractual data access, who holds the keys, whether configuration is reproducible — are all decisions taken long before, by people who were not thinking about incident response when they took them.
Scale, for context on how often this is exercised. ENISA's threat landscape report published on 1 October 2025 analysed 4,875 incidents over the twelve months from 1 July 2024. Cloud incidents are not a rare event class being planned for speculatively.
The rehearsal that is worth more than the plan. Take one real service and walk it through: an attacker authenticates as a legitimate administrator at 02:00 on a Sunday. Which alert fires, to whom, on what device? Which logs do you pull, and do you hold them? What do you ask the provider for, through which channel, and who is authorised to ask? What can you revoke yourself, immediately, without the provider? Nearly every organisation that runs that exercise finds the same two gaps: nobody knows the provider escalation route, and the logs stop short of where they need to start.
Backups that survive an attacker
Resilience in a cloud is usually discussed as availability — regions, zones, failover. The harder case is an adversary with valid credentials who wants your data gone, and it defeats most cloud backup designs because those designs assume failure, not malice.
The counter-intuitive warning worth reading twice. CISA's ransomware guidance, developed through the Joint Ransomware Task Force, states it directly: "Automated cloud backups may not be sufficient because if local files are encrypted by an attacker, these files will be synced to the cloud, possibly overwriting unaffected data." Synchronisation is not backup. A service that faithfully replicates your files will faithfully replicate their encryption, and a retention window shorter than your detection time means the good copies age out before anyone notices.
What the guidance actually asks for. "Maintain offline, encrypted backups of critical data, and regularly test the availability and integrity of backups in a disaster recovery scenario." Three requirements in one sentence, and the third is the one that fails: a backup never restored is a hypothesis. The guidance gives the reason offline matters in adversarial terms rather than technical ones — "most ransomware actors attempt to find and subsequently delete or encrypt accessible backups to make restoration impossible unless the ransom is paid." Accessible is the operative word. A backup reachable with the credentials an attacker already holds is part of the blast radius.
On immutability, with the caveat the source itself gives. The same guidance notes that some providers offer immutable storage that can protect data without a separate environment, and then adds a warning rarely repeated by the people selling it: "Use immutable storage with caution as it does not meet compliance criteria for certain regulations and misconfiguration can impose significant cost." Immutable storage is a strong control with two real edges — data you cannot delete is data you cannot delete when a regulator or a data subject requires you to, and a retention lock set wrongly is a bill you cannot stop. Use it deliberately, with the retention period decided by someone who has read both the threat model and the retention obligations.
The parts of a cloud estate people forget to back up. Not the data — everybody remembers the data. The guidance points at two others: keep offline copies of infrastructure-as-code template files "to quickly redeploy resources", and store applicable source code or executables with offline backups, noting that rebuilding from system images is more efficient but some images will not install correctly on different hardware or platforms. Add to that the configuration of the identity provider, the detection rules you wrote in lesson five, and the key material without which restored data is ciphertext.
The test that settles it. Assume an administrator identity is fully compromised and has been for three weeks. From that starting point, what can you restore, from where, using which credentials, and how long does it take? An answer that depends on anything the compromised identity could reach or delete is not an answer. Most organisations discover their recovery plan and their attack surface share a credential.
Concentration, assurance and the way out
The last lesson is the one that is not technical, and it is the one that senior people will ask you about. You cannot audit your provider. So on what basis do you rely on it, and what happens if you have to stop?
Why this became a regulatory question. Because supervisors noticed that a large part of the financial system now depends on a small number of providers. The EU's Digital Operational Resilience Act, which entered into application on 17 January 2025, creates an oversight framework whose stated purpose includes addressing "potential systemic and concentration risks arising from the financial sector's reliance on a limited number of ICT providers", alongside requirements on third-party risk management, monitoring providers and key contractual provisions. In the UK, the Prudential Regulation Authority's supervisory statement on outsourcing and third party risk management, first published in March 2021 and in force in its current form since 31 December 2024, has chapters that expand on European outsourcing expectations specifically in the areas of data security and of business continuity and exit plans — with a further version already published for 2027. Singapore's Monetary Authority has had technology risk management guidelines in place since January 2021. And the European Banking Authority finalised guidelines on third party risk management in September 2026 that, once applicable, will repeal its earlier outsourcing guidelines and focus on arrangements supporting critical or important functions.
The direction of travel, stated plainly. The frameworks are converging on three obligations, and you can prepare for all three without reading any of them in full: know which of your providers support a function you could not lose; have an exit that has been thought through rather than asserted; and be able to produce the register on demand.
What assurance you can actually get. Not an audit. What you can get is: third-party attestations and certifications, with their scope read rather than their logo noted; the provider's own published commitments; configuration you can verify yourself against an external baseline, as in lesson six; and your own telemetry, as in lesson five. The Cyber Safety Review Board's recommendation that providers "adopt incident and vulnerability disclosure practices to maximize transparency across and between their customers, stakeholders, and the United States government, even in the absence of a regulatory obligation to report" is a useful procurement question in its own right: what does this provider commit to telling us, and has it done so historically?
On exit, honestly. A documented exit plan for a deeply integrated managed service is often a plan to rebuild, and the realistic output of the exercise is usually not a migration runbook but a costed, dated answer to "how long would this take and what would it cost" — which is information a board can act on. The plan that says "we would migrate to another provider" without that number has not been written.
Where this course leaves you. Nothing here resolves the central asymmetry: you depend on controls you cannot inspect, operated by an organisation whose own regulator-commissioned review found it could not explain how its most sensitive key was taken. What you can do is make your own side of the line genuinely strong — identity, keys, logging, baselines, segmentation, workloads, recovery — and know, in writing and in numbers, what you would do if the other side failed. That is not a lesser goal. On the evidence of the one incident in this course that was investigated in public, it is the goal that worked.
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 800-207: Zero Trust Architecture, 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 800-190: Application Container Security Guide, NIST SP 800-61 Revision 3: Incident Response Recommendations and Considerations for Cybersecurity Risk Management
- MITRE ATT&CK — ATT&CK cloud matrix
- ENISA — ENISA Threat Landscape 2025
National detail, by market
This course teaches the international standard and one worked example. The national detail for a market sits in its own pack, so it can be kept current without reopening the course: India.
Sources
The official documents this course relies on. Laws and guidance change, so check the current version.
- Review of the Summer 2023 Microsoft Exchange Online Intrusion · Cyber Safety Review Board, US Department of Homeland Security
- Cloud security guidance · UK National Cyber Security Centre
- Cloud security principle 14: secure use of the service · UK National Cyber Security Centre
- Lightweight approach to cloud security · UK National Cyber Security Centre
- NIST SP 800-207: Zero Trust Architecture · 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 800-190: Application Container Security Guide · 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
- Cloud Security Technical Reference Architecture, version 2 · CISA, United States Digital Service and FedRAMP
- Secure Cloud Business Applications (SCuBA) project · US Cybersecurity and Infrastructure Security Agency
- #StopRansomware Guide · US Cybersecurity and Infrastructure Security Agency
- ATT&CK cloud matrix · The MITRE Corporation
- ENISA Threat Landscape 2025 · European Union Agency for Cybersecurity
- CIS Benchmarks · Center for Internet Security
- SS2/21: Outsourcing and third party risk management · Bank of England, Prudential Regulation Authority
- Digital Operational Resilience Act (DORA) · European Insurance and Occupational Pensions Authority
- Guidelines on third party risk management · European Banking Authority
- Guidelines on Risk Management Practices – Technology Risk · Monetary Authority of Singapore



