The Complete Guide to Cloud Security
How cloud breaches actually happen, what the shared responsibility model covers, and the identity, storage and logging controls that matter most.

A developer needs to share a database export with a contractor on Friday afternoon. The quickest route is a storage bucket, so they create one, upload the file, and switch off the setting that was blocking public access because the contractor could not download it otherwise. The contractor gets the file. Nobody switches the setting back. Eighteen months later the bucket still holds that export plus forty others, and it has been catalogued by one of the services that continuously scan for open cloud storage.
No vulnerability was exploited. No password was stolen. The provider's platform worked exactly as designed. This is what the majority of cloud security incidents look like, and it is why cloud security is less about attacks on the provider and more about the decisions made by the people using it.
This guide covers how responsibility is actually divided between you and your provider, why configuration is the dominant failure mode, how cloud identity works and how attackers abuse it, where data gets exposed, what to log, and how to handle more than one platform. It assumes you use cloud services at work and need to understand the real risks rather than a vendor's version of them.
What the cloud actually changes
NIST's definition of cloud computing describes five characteristics: on-demand self-service, broad network access, resource pooling, rapid elasticity and measured service. Two of those create most of the security consequences.
On-demand self-service means anyone with the right permissions can create infrastructure in seconds without asking anyone. The approval step that used to exist because somebody had to physically install a server has gone.
Broad network access means what you create is reachable over the network by default unless you actively arrange otherwise. In a traditional data center a misconfigured server was wrong but usually still sat behind a perimeter firewall. In cloud, a misconfigured service can be wrong and immediately reachable from anywhere.
Put those together and you get the central shift: the consequences of a small mistake arrive faster and travel further. The compensating advantage is that everything is an API call, so configuration can be inspected, tested and corrected programmatically in a way physical infrastructure never allowed.
The shared responsibility model
Every major provider publishes a version of the same diagram. AWS splits it into security "of" the cloud versus security "in" the cloud. Microsoft publishes a responsibility table by service type. Google frames it as shared responsibility moving toward "shared fate", where the provider takes a more active role in helping customers configure things safely.
The useful way to read all of them is as a sliding line. The more of the stack you rent, the less of it you configure, and the less there is for you to get wrong.
| Layer | Infrastructure as a service | Platform as a service | Software as a service |
|---|---|---|---|
| Facility, hardware, hypervisor | Provider | Provider | Provider |
| Operating system and its patching | You | Provider | Provider |
| Runtime and middleware | You | Provider | Provider |
| Application code and its flaws | You | You | Provider |
| Network and firewall configuration | You | Shared | Provider |
| Identities, roles and permissions | You | You | You |
| Data, its classification and its access | You | You | You |
| Configuration of the service itself | You | You | You |
Notice what never moves. Identity, data and configuration are yours in every model, including the most managed product you buy. A provider can hand you a perfectly patched platform and you can still give an attacker a global administrator account.
Where the model misleads people
The diagrams are accurate and still cause three recurring mistakes.
The first is treating "the provider handles security" as a conclusion rather than a boundary. Reading the diagram as reassurance inverts its purpose. It exists to tell you precisely what is not covered.
The second is assuming the line sits in the same place across services from the same provider. It does not. A managed database service patches the database engine for you; a virtual machine running the same engine does not. Two services in one account can sit on different sides of the line, and the provider's documentation for that specific service is the only reliable guide.
The third is believing the model transfers accountability. It divides technical work, not consequences. If customer data leaks because of a setting you chose inside a vendor's product, your customers, your regulator and your insurer all come to you. Cyber insurance policies and supplier contracts rarely shift that back, and both tend to assume you configured the service competently.
Misconfiguration is the dominant cause
When NSA and CISA red and blue teams published their assessment of the most common misconfigurations found across large networks, the list was not exotic. It included default configurations of software and applications, improper separation of user and administrator privilege, insufficient internal network monitoring, poor patch management, bypass of system access controls, weak or misconfigured multi-factor authentication methods, insufficient access control lists on network shares and services, poor credential hygiene, and unrestricted code execution. CISA's joint advisory on weak controls routinely exploited for initial access makes the same point about how intrusions begin: multi-factor authentication not enforced, incorrectly applied privileges, software not up to date, default credentials left in place, and services exposed that should not be.
None of that is unique to cloud, but cloud amplifies all of it, for structural reasons.
- Defaults optimize for things working. A new service that refuses connections generates support tickets, so defaults tend to be permissive enough to function. Several providers have tightened this considerably, but the pattern persists across the thousands of settings in a large account.
- The surface is enormous. A single major provider offers hundreds of services, each with its own permission model, networking behavior and logging configuration. No one person understands all of them.
- Infrastructure is created by developers. The person creating a queue, a function and a storage bucket at 11pm is solving a product problem, not performing a security review.
- Temporary changes become permanent. The "just for today" relaxation of a control has no expiry date and no reminder attached.
- Copy and paste propagates errors. An over-permissive policy written once gets reused across twenty services because it works.
The practical response is to stop relying on individuals remembering, and to make the safe state the default state. Account-wide and organization-wide guardrails applied centrally are worth more than any amount of training on individual settings. Block public storage access at the account level so a single bucket setting cannot override it. Use organization policies to deny the regions, services and configurations you have decided against. Scan infrastructure-as-code before it deploys rather than auditing what already exists.
Identity is the real perimeter
In a traditional network an attacker needed network position. In cloud they need credentials. Everything is an authenticated API call, which makes identity and access management the control that decides what an intrusion becomes.
MITRE ATT&CK maintains a Cloud matrix documenting the techniques adversaries use against cloud environments, and reading through it is instructive. The entries are dominated by credential access, account manipulation and permission abuse rather than exploitation of platform flaws.
The identity failures that matter
Long-lived access keys. A static key pair with no expiry is the most common cloud credential leak. Keys end up in source repositories, container images, build logs, support tickets and laptop configuration files. Automated scanners find published keys quickly. The fix is to stop issuing them: use workload identities and role assumption so running code receives short-lived credentials from the platform, and reserve static keys for the small number of cases that genuinely cannot work any other way.
Permissions granted with a wildcard. A policy allowing every action on every resource is quick to write and impossible to reason about. The realistic approach is to start permissive during development and then tighten using the access data the platform already collects: every major provider can report which permissions an identity has actually used, which turns least privilege from a guessing exercise into a cleanup task.
Missing or weak multi-factor authentication. Enforcing a second factor on human administrative identities is the highest-value change available, and the method matters. Push-approval prompts can be worn down by repeated requests until someone taps approve, and codes sent by text can be intercepted through SIM swapping. Hardware security keys and passkeys resist phishing in a way the others do not. Our guide to enforcing two-step verification across a team covers the rollout.
Identities that can change permissions. An identity able to modify policies, create roles or attach itself to a more privileged role is effectively an administrator regardless of what its own permissions currently say. Escalation paths through permission-modifying rights are among the least understood parts of cloud IAM and among the most reliably abused. Enumerate who can grant permissions and treat that list as your true administrator list.
Federation and third-party access. Identity providers, build platforms, monitoring vendors and backup tools all hold roles in your environment, and each is a path in. Scope them tightly, require conditions on which external identities may assume which roles, and review them when a tool is retired. Attacker-controlled OAuth applications are a related problem; see consent phishing, where the attacker asks a user to approve an application rather than hand over a password.
The root and owner accounts
Every cloud account has an original identity that cannot be fully constrained by the normal permission system. It should have a hardware second factor, no access keys, a password stored so that two people must act together to retrieve it, an alert on every use, and no routine role at all. Day-to-day administration belongs to separate identities that can be revoked without drama.
Storage and data exposure
Open object storage is the archetypal cloud leak because the failure is silent. Nothing breaks when a bucket becomes public. There is no error, no alert by default, and no sign of it on the application side.
Providers have added layered defenses worth using properly. Account-level public access blocks override individual bucket settings, so an engineer cannot make a bucket public by accident even by changing the bucket's own configuration. Ownership controls can disable the older per-object access lists entirely, removing a whole category of confusing interactions. Pre-signed or time-limited URLs let you share a single file with someone outside the organization without making anything public, which is the control that would have solved the Friday afternoon problem at the top of this guide.
Databases and search indexes fail the same way but worse, because some ship with authentication optional. A managed database placed on a public subnet with a permissive security group and a weak password is reachable by anyone running a port scan, and the first indication of trouble is often an extortion note left in a renamed table.
Three habits prevent most of this.
- Default to private and share by exception, with the exception implemented as a time-limited link rather than a configuration change.
- Know where the sensitive data is. You cannot protect data you have not located. Discovery tooling is imperfect but better than assumption, and a simple inventory of which stores hold personal or financial data is enough to start.
- Separate environments by account, not by naming convention. Production data in a development account is protected by development's controls. Account boundaries are the strongest isolation the platform offers; prefixes in resource names are not isolation at all.
On encryption, be clear about what it buys. Provider-managed encryption at rest is worth enabling everywhere and defeats one specific threat: someone obtaining the physical storage media. It does nothing against an attacker holding valid credentials, because the platform decrypts transparently for any authorized caller. Customer-managed keys add value mainly because they let you revoke access and because key usage is logged, giving you a second place to detect unusual access. Treating encryption as a substitute for access control is the mistake to avoid.
Logging and knowing what happened
The question that decides how a cloud incident goes is simple: can you reconstruct what the attacker did? Most organizations discover the answer during the incident.
Cloud logs come in distinct layers, and they are not equally useful.
- Control plane and audit logs record API calls: who created, modified, read or deleted what. These are the most valuable logs you have, because almost every attacker action in cloud is an API call.
- Identity and sign-in logs record authentication events, including the failures and unusual locations that often precede a successful intrusion.
- Data access logs record reads of individual objects. Often off by default, often expensive, and the only way to answer the question regulators ask: what was actually accessed?
- Network flow logs record connections. Useful for scoping lateral movement and for noticing exfiltration volume.
- Application and operating system logs remain entirely your responsibility in infrastructure services.
Four decisions matter more than the choice of tooling. Turn on audit logging across every account and region, including the ones nobody uses, because unused regions are where unmonitored activity goes. Extend retention beyond the default, which is frequently measured in weeks while intrusions routinely go undetected for months. Send a copy to a separate account whose credentials your production identities do not hold, so an attacker who gains administrative access cannot erase the record of it; disabling or deleting logs is a documented and common adversary behavior. Finally, alert on a short list of events that should essentially never happen: root account use, changes to logging configuration, new identity providers or federation trusts, policy changes granting broad permissions, and mass deletion.
For business SaaS platforms rather than infrastructure, CISA's Secure Cloud Business Applications project publishes concrete configuration baselines for widely used cloud business services, and its playbook on expanded cloud logs covers how to get investigative value out of the logging tiers available in Microsoft's cloud. Both are a reasonable external reference for what "configured properly" means.
Multi-cloud, hybrid and SaaS sprawl
Multi-cloud is often presented as resilience. Security-wise it usually is not. Each additional platform brings a different permission model, a different logging format, a different set of defaults and another console your team must know well. Depth of knowledge in one platform prevents more incidents than breadth across three.
Deliberate multi-cloud, where a specific workload runs on a specific platform for a clear reason, is manageable. Accidental multi-cloud, where three teams each chose a provider and nobody can list what exists, is the common case and the dangerous one.
The same problem appears more acutely with SaaS. A typical organization has more business-critical data in SaaS products than in infrastructure it configured itself, and far less visibility into it. Each product has its own administrator model, sharing settings, log retention and integration permissions. A minimum standard is worth defining and applying to every SaaS product you adopt: single sign-on through your identity provider so access ends when employment does, enforced multi-factor authentication, administrator accounts named and reviewed, external sharing restricted, logs exported somewhere you control, and a documented owner. Routing every product through single sign-on is what makes the leaver process work, because otherwise each product keeps its own copy of an account you have forgotten about.
Hybrid setups, where cloud identity is linked to an on-premises directory, deserve specific attention because the link works in both directions. Compromise of the on-premises directory can be used to reach cloud resources, and compromise of a cloud identity holding directory synchronization rights can be used to reach the internal network. Treat the synchronization servers and service accounts as among the most sensitive systems you operate.
Where cloud incident response differs
Cloud changes the mechanics of response in ways worth deciding in advance rather than during an incident.
Containment is an API call, which is fast and dangerous. Revoking a role, rotating a key or detaching a policy can stop an attacker in seconds and can take down production in the same second. Know which identities are safe to disable immediately and which are not.
Evidence is ephemeral. A compromised container or virtual machine may be replaced by automatic scaling before anyone looks at it, taking the evidence with it. Snapshot before you terminate, and check that your platform's auto-remediation does not destroy the thing you need to examine.
Scope is a permissions question. The first question is not "which machine" but "what could that identity reach?" The answer comes from the permission policies, not from the network diagram.
Provider support is part of your plan. Know which support tier you hold, what the provider will and will not investigate on your behalf, and which logs only they have. Find this out while nothing is wrong.
A realistic order of work
If you are improving an existing cloud environment rather than designing a new one, sequence matters. This order reflects what prevents the most harm per hour spent.
- Inventory. List every cloud account, subscription and SaaS product, with a named owner and what it holds. You cannot secure an unknown account, and unknown accounts are where the worst findings live.
- Lock down administrative identity. Hardware second factors for administrators, root and owner identities taken out of daily use, static keys inventoried with a date set for removing each one.
- Turn on logging properly. All accounts, all regions, extended retention, a copy somewhere separate, and alerts on the handful of events that should never happen.
- Close the exposure. Public storage blocked at account level, databases off public subnets, management interfaces not reachable from the internet, and a review of what your provider's posture tooling already flags for free.
- Set guardrails. Organization-level policies that deny the configurations you have decided against, so individual mistakes cannot create them.
- Reduce permissions using evidence. Use the platform's own access analysis to remove unused permissions, starting with identities that can grant permissions to others.
- Then buy tooling, if you still have a visibility gap you can describe precisely. Tools bought before this point mostly generate findings nobody acts on.
Related work sits alongside this. Vulnerability management covers the patching and prioritization side, zero trust covers the access architecture this points toward, and the small business cybersecurity guide covers the broader program if you are starting from very little.
Mistakes worth avoiding
- Assuming the provider's certifications cover you. Their compliance report describes their controls, not your configuration.
- Lifting and shifting an internal network into cloud unchanged. A flat internal network becomes a flat internet-adjacent network, and the perimeter controls that made it tolerable are no longer there.
- Leaving development and test accounts unmanaged. They hold real data more often than anyone admits, and they are a frequent entry point.
- Treating credentials in a private repository as safe. Repository visibility changes, forks happen, and laptops get stolen.
- Measuring progress by the number of findings closed. Count what is exposed and who holds administrative power instead.
Not sure whether a message about your cloud account is genuine? Fake provider alerts and billing warnings are a common way in. Paste the message or link into CyberWatch AI Scan for a free check before you click anything.
The short version
Cloud providers run the platform well. The failures that matter happen above their line, in identity, configuration and data handling, and they are rarely sophisticated. An attacker finds a leaked key or phishes an administrator who has no hardware second factor, discovers that the identity holds far more permission than its job requires, reads data from a store nobody knew held anything sensitive, and leaves little trace because logging was off or had already expired.
Every step in that sequence is preventable with controls the provider already gives you. The work is unglamorous: know what you have, constrain who can change it, record what happens, and make the safe configuration the one that applies unless somebody deliberately overrides it.
Frequently asked questions
What is the shared responsibility model?
It is the split between what your cloud provider secures and what you secure. The provider is responsible for the security of the cloud itself, meaning physical facilities, hardware and the virtualization layer. You are responsible for security in the cloud: your identities, permissions, configurations, data and code. The exact dividing line moves depending on whether you buy infrastructure, a platform or a finished application.
What causes most cloud breaches?
Configuration and identity problems far more often than flaws in the provider's platform. Storage left open to the internet, over-permissive roles, long-lived access keys that leak, missing multi-factor authentication on administrator accounts and disabled logging account for a large share of real incidents.
Is the cloud more secure than my own servers?
The underlying platform is usually maintained better than a small organization could manage alone, because patching the hypervisor and securing the data center is the provider's full-time job. Everything above that line is still yours, and cloud makes mistakes faster to make and easier to reach from anywhere.
What is the single most useful cloud security control?
Strong, phishing-resistant multi-factor authentication on every administrative identity, combined with removing long-lived access keys in favor of short-lived credentials. Identity is how most cloud intrusions begin and how they escalate.
Do I need a cloud security product?
Not at the start. The provider's own guardrails, such as account-level public access blocks on storage, organization policies, centralized logging and built-in posture tooling, cover the common failure modes. Buy tooling once you know what you are trying to see and cannot see it.
How long should I keep cloud logs?
Long enough to investigate an intrusion you discover late. Control plane and identity logs are the most valuable, and default retention is often far shorter than the time it takes to notice a compromise, so extend it deliberately and store a copy somewhere the same credentials cannot delete.
Is multi-cloud more secure than one provider?
Usually less secure in practice. It removes a single point of failure but multiplies the number of permission models, logging formats and consoles your team must understand well. Most organizations get better security from knowing one platform deeply.
Who is responsible if a SaaS provider is breached?
The provider is responsible for their platform, but the consequences land on you, and your contract rarely transfers much. In practice you remain accountable for the configuration choices you made in their product, the data you put there, the accounts you granted and your own notification duties.
What should I check first in a new cloud account?
Multi-factor authentication and no long-lived keys on the root or owner identity, public access blocked at the account level for storage, logging enabled and sent somewhere separate, billing alerts switched on, and a review of every identity that can create or change permissions.
Does encryption protect cloud data from attackers?
Provider-managed encryption at rest protects against someone taking the physical disks, which is not how cloud data is usually stolen. It does nothing against an attacker using valid credentials, because the platform decrypts the data for them transparently. Access control is the control that matters for that threat.
Sources
- SP 800-145, The NIST Definition of Cloud Computing, US National Institute of Standards and Technology
- NSA and CISA Red and Blue Teams Share Top Ten Cybersecurity Misconfigurations, US Cybersecurity and Infrastructure Security Agency
- Weak Security Controls and Practices Routinely Exploited for Initial Access, US Cybersecurity and Infrastructure Security Agency
- Shared Responsibility Model, Amazon Web Services
- Shared responsibility in the cloud, Microsoft Learn
- Shared responsibilities and shared fate on Google Cloud, Google Cloud Architecture Center
- Blocking public access to your Amazon S3 storage, Amazon Web Services Documentation
- Cloud security guidance, UK National Cyber Security Centre
- Cloud Matrix, MITRE ATT&CK
- Secure Cloud Business Applications (SCuBA) Project, US Cybersecurity and Infrastructure Security Agency


