This is part two of a series. You can read part one here or track the project on GitHub.

How the Cloud Is Different for Security

In the early days of cloud computing, even some very well-respected security professionals claimed it was little more than a different kind of outsourcing, or equivalent to the multitenancy of a mainframe. But the differences run far deeper, and we will show how they require different cloud security controls. We know how to manage the risks of outsourcing or multi-user environments; cloud computing security builds on this foundation and adds new twists.

These differences boil down to abstraction and automation, which separate cloud computing from basic virtualization and other well-understood technologies.


Abstraction is the extensive use of multiple virtualization technologies to separate compute, network, storage, information, and application resources from the underlying physical infrastructure. In cloud computing we use this to convert physical infrastructure into a resource pool that is sliced, diced, provisioned, deprovisioned, and configured on demand, using the automation we will talk about next.

It really is a bit like the matrix. Individual servers run little more than a hypervisor with connectivity software to link them into the cloud, and the rest is managed by the cloud controller. Virtual networks overlay the physical network, with dynamic configuration of routing at all levels. Storage hardware is similarly pooled, virtualized, and then managed by the cloud control layers. The entire physical infrastructure, less some dedicated management components, becomes a collection of resource pools. Servers, applications, and everything else runs on top of the virtualized environment.

Abstraction impacts security significantly in four ways:

  • Resource pools are managed using standard, web-based (REST) Application Programming Interfaces (APIs). The infrastructure is managed with network-enabled software at a fundamental level.
  • Security can lose visibility into the infrastructure. On the network we can’t rely on physical routing for traffic inspection or management. We don’t necessarily know which hard drives hold which data.
  • Everything is virtualized and portable. Entire servers can migrate to new physical systems with a few API calls or a click on a web page.
  • We gain greater pervasive visibility into the infrastructure configuration itself. If the cloud controller doesn’t know about a server it cannot function. We can map the complete environment with those API calls.

We have focused on Infrastructure as a Service, but the same issues apply to Platform and Software as a Service, except they often offer even less visibility.


Virtualization has existed for a long time. The real power cloud computing adds is automation. In basic virtualization and virtual data centers we still rely on administrators to manually provision and manage our virtual machines, networks, and storage. Cloud computing turns these tasks over to the cloud controller to coordinate all these pieces (and more) using orchestration.

Users ask for resources via web page or API call, such as a new server with 1tb of storage on a particular subnet, and the cloud determines how best to provision it from the resource pool; then it handles installation, configuration, and coordinating all the networking, storage, and compute resources to pull everything together into a functional and accessible server. No human administrator required.

Or the cloud can monitor demand on a cluster and add and remove fully load-balanced and configured systems based on rules, such as average system utilization over a specified threshold. Need more resources? Add virtual servers. Systems underutilized? Drop them back into the resource pool. In public cloud computing this keeps costs down as you expand and contract based on what you need. In private clouds it frees resources for other projects and requirements, but you still need a shared resource pool to handle overall demand. But you are no longer stuck with under-utilized physical boxes in one corner of your data center and inadequate capacity in another.

The same applies to platforms (including databases or application servers) and software; you can expand and contract database storage, software application server capacity, and storage as needed – without additional capital investment.

In the real world it isn’t always so clean. Heavy use of public cloud may exceed the costs of owning your own infrastructure. Managing your own private cloud is no small task, and is ripe with pitfalls. And abstraction does reduce performance at certain levels, at least for now. But with the right planning, and as the technology continues to evolve, the business advantages are undeniable.

The NIST model of cloud computing is the best framework for understanding the cloud. It consists of five Essential Characteristics, three Service Models (IaaS, PaaS, and SaaS) and four Delivery Models (public, private, hybrid and community). Our characteristic of abstraction generally maps to resource pooling and broad network access, while automation maps to on-demand self service, measured service, and rapid elasticity. We aren’t proposing a different model, just overlaying the NIST model to better describe things in terms of security.

Thanks to this automation and orchestration of resource pools, clouds are incredibly elastic, dynamic, agile, and resilient.

But even more transformative is the capability for applications to manage their own infrastructure because everything is now programmable. The lines between development and operations blur, offering incredible levels of agility and resilience, which is one of the concepts underpinning the DevOps movement. But of course done improperly it can be disastrous.

Cloud, DevOps, and Security in Practice: Examples

Here are a few examples that highlight the impact of abstraction and automation on security. We will address the security issues later in this paper.

  • Autoscaling: As mentioned above, many IaaS providers support autoscaling. A monitoring tool watches server load and other variables. When the average load of virtual machines exceeds a configurable threshold, new instances are launched from the same base image with advanced initialization scripts. These scripts can automatically configure all aspects of the server, pulling metadata from the cloud or a configuration management server. Advanced tools can configure entire application stacks. But these servers may only exist for a short period, perhaps never during a vulnerability assessment window. Or images may launch in the wrong zone, with the wrong network security rules. The images and initialization scripts might not be up to date for the latest security vulnerabilities, creating cracks in your defenses.
  • Immutable Servers: Autoscaling can spontaneously and automatically orchestrate the addition and subtraction of servers and other resources. The same concepts can eliminate the need to patch. Instead of patching a running server you might use the same scripting and configuration management techniques, behind a virtual load balancer, to launch new, up-to-date versions of a server and then destroy the unpatched virtual machines.
  • Snapshots: Cloud data typically relies on virtual storage, and even running servers use what are essentially virtual hard drives with RAID. A snapshot is a near-instant backup of all the data on a storage volume, since it merely needs to copy one version of the data, without taking the system down on affecting performance. These snapshots are incredibly portable and, in public clouds, can be made public with a single API call. Or you could write a program to snapshot all your servers at once (if your cloud has the capacity). This is great for forensics, but also enables an attacker to copy your entire data center and make it public with about 10 lines of code.
  • Management Credentials: The entire infrastructure deployed on the cloud is managed, even down to the network and server level, using API calls and perhaps web interfaces. Administrator tools typically keep these credentials in memory as environment variables or the registry, making them accessible even without administrative control over the cloud admin’s workstation. Also, most clouds don’t provide an audit log of these commands. Many organizations fail to compartmentalize the rights of cloud administrators, leaving their entire infrastructure open to a single compromised system.
  • Software Defined Security: With only 20 lines of code you can connect to your cloud over one API, your configuration management tool with another, and your security tool with a third. You can instantly assess the configuration and security of every server in your environment, without any scanning, in real time. This is nearly impossible with traditional security tools.

Snapshots highlights some of the risks of abstraction. Autoscaling, some risks of automation, and management credentials, the risks of both. But Software Defined Security and immutable servers offer advantages. We will dig into specifics next, now that we have highlighted the core differences.

And all that without mentioning multitenancy or outsourcing.