We’ve recently seen a spate of news stories about hardware-based attacks. For instance, two recent attempted bank heists at Bank Santander and Barclay’s involved criminals stealing millions of dollars via malicious hardware devices. More concerning, recently leaked documents indicate that the NSA may have collaborated with hardware manufacturers to subvert cryptographic hardware implementations. Researchers recently proposed new ways to create hardware backdoors at the sub-gate level, making it hard to detect even to someone inspecting circuit layouts. But are these hardware risks relevant to servers that we use in the cloud?
Modern servers are comprised of many components: processors (CPUs), memory (RAM), disks, buses, network cards, and human interface devices. Each of these components pass through the hands of manufacturers, vendors, supply chains, integrators, and service personnel before ultimately ending up inside servers processing your sensitive data. That server itself may be housed off-premise or be leased from another organization, such as a cloud service provider. Most organization rely on at least some servers that are outside their physical control.
The risk with this loss of control is that is that anyone with access to those components, at any moment in time, could compromise a component or substitute a malicious device in its place. There are many well known boot integrity vector attacks where an attacker could subvert firmware in a system, for example the “Evil Maid Attack”.
Network cards are particularly risky since they have Direct Memory Access (DMA) to all system memory and can exfiltrate stolen data over the network. Some network devices have been found to have remotely exploitable vulnerabilities that allow an attacker to take control of the card and subsequently the host system.
Enterprises are increasingly adopting in-memory architectures to reduce application latency. However, as in-memory architectures become more common, more sensitive data is persisted in memory for long periods of time. While commonly used RAM is generally volatile, it can actually persist its contents after a system loses power. This allows an attacker to literally freeze memory and read its contents in what’s called a “cold boot” attack.
More worrisome, RAM is becoming more persistent with technologies like non-volatile memory in DDR3 form factors. By design, these memory technologies persist contents of memory when power is lost — just like a disk. Attackers could simply walk away with a memory DIMM containing not only private data and software, but also critical cryptographic keys used to secure data at-rest encryption.
PrivateCore’s philosophy to addressing vulnerabilities in hardware and the risk of persistent memory is to minimize the number of components that users must trust in a server. With today’s technology, it’s possible to reduce a server’s security boundary to just one component: the CPU. From just the CPU, it’s possible to establish a trusted compute base safe from the rest of the components in the system.
Protecting against hardware threats is not easy. Some organizations closely audit their supply chains to ensure the provenance of firmware and devices in their systems, and operate their servers in tightly controlled physical environments. For cloud environments or remote locations, this may not be an option. In those cases, minimizing the trust perimeter to a single component may be the best option to reduce the threat of hardware vulnerabilities and protect sensitive data in-use.