Suppose a public urban transport operator that installs its online ticket sales system with a cloud provider. The contract is signed, the first virtual servers are running, and the chief information officer asks the team: how do you secure a cloud infrastructure when the operator owns neither the server rooms nor the machines?
The method we propose comes down to six steps, in this order: establish what the provider protects and what it leaves to the customer, settle identities and access, fix the configuration of resources, decide who holds the encryption keys, organise logging, then back up and prove restoration. Each step leaves a verifiable deliverable and names a person who answers for it. The order follows the dependencies: the split says what remains to be handled, and access, who can undo the rest.
The scope stops at the infrastructure: application code is covered in our article on application security, and what to do during an attack in the one on cyber crisis management.
To secure a cloud infrastructure, first establish what the provider leaves to the customer
A cloud provider secures what it operates; the customer answers for what it places there and how it configures it. Providers such as Amazon Web Services or Microsoft Azure publish this split under the name of shared responsibility model and specify that it varies according to the service used.
| Type of service | The provider handles | The customer retains |
|---|---|---|
| Infrastructure (IaaS) | Data centres, network and physical machines, virtualisation layer | Operating system and its patches, network rules, applications, identities, configuration, data |
| Platform (PaaS) | In addition: operating system and runtime environment | Service settings, deployed code, identities, data |
| Software (SaaS) | In addition: the application itself | Configuration, user accounts and rights, data |
This table gives an overall reading: the exact boundary is read in the documentation and contract of each service. A qualification under SecNumCloud, ANSSI’s framework for cloud service providers, covers a service of the provider, and therefore what the provider operates. The right-hand column remains intact. The level of assurance to require from the provider is chosen according to the data hosted, as set out on our page on sovereign cloud.
This step produces a matrix: for each service subscribed to, what the provider guarantees and the name of the person who answers for the rest. The chief information security officer (CISO) keeps it with the cloud architect.
Identities and access before the first resource
In a cloud, an account with extensive rights can create, expose or delete resources through a call to the provider’s administration interface. Access to this interface is therefore settled before anything else.
Administrators authenticate through the organisation’s directory, with a second factor. Rights are granted by role, on a least-privilege basis, and for a limited period when they are elevated. Sealed emergency accounts cover the case where the directory goes down, and each use triggers an alert. Machines and deployment pipelines receive short-lived credentials, and no secret appears in the code.
For the entities it covers, Directive (EU) 2022/2555, known as NIS2, cites in Article 21(2) access control policies (point (i)) and multi-factor authentication, “where appropriate” (point (j)). Its reach in France depends on transposition. For financial entities, Article 9(4)(c) of Regulation (EU) 2022/2554, known as DORA, limits access to assets to what is required for legitimate and approved functions and activities.
The role model, the named list of privileged accounts and the emergency account procedure are entrusted to the identity management owner. Each application owner periodically confirms the rights open on their scope.
A configuration described as code, then monitored
A storage bucket readable by everyone, a database reachable from the internet, an administration port left open: these configuration errors are the customer’s responsibility.
The platform team, in the sense of our DevOps and platform engineering page, carries out this step based on rules set by the CISO. It first lays down the structure: accounts separated by environment and by sensitivity, a segmented network, no public exposure by default. It adds preventive rules, applied across the whole organisation, which refuse the prohibited action, for example creating a resource outside the authorised regions. Finally, it continuously compares the actual state of resources with the desired state.
The desired state is written with an infrastructure as code tool such as Terraform: each change is reviewed before being applied and leaves a trace. In IaaS, operating system patches fall within the scope, since the provider does not apply them. The deliverable is twofold: the versioned foundation, and the record of open deviations with their age.
Encrypting is a setting, holding the keys is a decision
Activating encryption of stored data and of exchanges is, on a cloud service, a matter of configuration. The decision concerns the keys. They can be left with the provider. The customer can administer them itself in the provider’s key management service, and thus decide on their rotation and revocation. It can also house them in a device it operates outside the provider.
Each degree of control adds an operating burden: a lost key makes the data it protects unreadable, along with the backups encrypted with it. It is better if the person who administers the keys is not the one who administers the data.
The CISO signs off the deliverable: one rule per data class, naming the option chosen, the rotation frequency and the roles authorised to use the key, supplemented by an inventory of keys. Choosing the class is a matter for the data owner.
Logs out of reach of those they monitor
The logs that matter for the infrastructure record administrative actions, data access and network flows. Their value as evidence collapses if a production administrator can delete them. Collection therefore ends up in a separate account, where logs are appended without being modifiable.
A few events justify an immediate alert: creation of a privileged account, logging being stopped, a resource made public, use of an emergency account. Each alert has a recipient and an instruction. The list of enabled sources, the retention periods and the alert table belong to the team in charge of detection.
Back up, then prove restoration
The availability promised by a provider does not replace a backup: our page on cyber resilience explains why, and how business units set recovery objectives.
The scope to be backed up goes beyond data. Infrastructure code, secrets and keys are part of what you need to be able to recover, failing which intact data remain unusable. Copies reside in a separate account, under different access rights, and their deletion is blocked for a period set in advance.
The proof is a test: rebuild an empty environment from the code, restore the data into it, measure the duration and compare it with the objective approved by the business units. For financial entities, Article 12 of DORA requires periodic testing of backup and restoration procedures. The head of production keeps the backup plan and the dated report of the last test.
These six deliverables are only valid when dated. The operator in our scenario will reread them for each new service subscribed to with its provider, starting again from the matrix: it is the matrix that says what the other five must cover.