Key Takeaways:
- With AI-driven demand for data center capacity making data center security more crucial than ever, the convergence of physical and cybersecurity can no longer be an afterthought.
- The boundaries of trust must expand to cover the entire supply chain, requiring new systems of defense collaboration.
- New security standards for data center components are driving component design but could also create additional bottlenecks.
Securing the entire data center supply chain is a monumental task, and the stakes keep rising as the value of data stored in hyperscale facilities increases by the minute.
Security developers are contending with rapidly scaling demand for data center capacity, unprecedented frontier cybersecurity threats, looming regulatory security deadlines, and the constant pressure to get their products to market and their clients’ facilities switched on as quickly as possible. To meet that demand, they’re designing systems that can continually verify the trustworthiness of the many moving parts that make up a modern data center.
Ensuring hardware’s cryptographic agility, managing the overlapping attack surfaces between a data center’s physical and cybersecurity, and tracking each component’s provenance across the global supply chain are all necessary parts of building a safe, secure modern facility. Developers are now finding ways to deliver products with more built-in complexity to meet these security demands.
The path to zero trust
Data center security developers are moving away from building security controls to block threats, focusing instead on designing “zero trust” systems that verify every step of the way.
With extremely sensitive information at risk, any hiccup in the supply chain can cause problems. “One of our customers in the Middle East trashed a huge order of servers because they traced the logistics of the shipment, and there was one period of 48 hours where no one could attest to the security of all those boxes. So they scrapped them,” said Scott Register, vice president of enterprise at Keysight Technologies. “The way we think about cybersecurity in the era of AI — even securing the data centers themselves outside of the models — is fundamentally different. Historically, we thought about controls, antivirus, or firewalls. Now we’re moving from thinking about controls to thinking about trust. If security is having the controls in place, then trust is having confidence backed by evidence. If you think about different domains of trust, you start with the supply chain.”
This shift in security design comes in direct response to attack strategies threatening data centers and other key infrastructure, including AI-aided remote physical attacks. “The evolution of remote physical attacks can be only mitigated by a paradigm shift in hardware-based security,” said Peter Laackmann, security expert at Infineon Technologies. “Conventional approaches for security domain isolation reach their limits and will be widely replaced by dynamic boundaries — for example, by locking digitally signed and verified identities to software images and dynamic virtual machines. Company networks are expected to evolve toward zero-trust architectures, away from traditional firewall concepts. A zero-trust approach means treating every access as if it’s from an untrusted communication partner and protecting it with authentication and encryption by default. State-of-the-art security components like certified secure elements and security-hardened MCUs build an ideal symbiosis to enable the three elements of information security crucial for data centers — confidentiality, integrity, and availability.”
The challenge now for security developers is to continually verify that every component within a massive data center is exactly what it’s supposed to be, checking for any exploitation that could have occurred across a vast international supply chain. “The supply chain is all about the time to market, and in this respect it’s clearly one avenue for subversion,” said Sylvain Guilley, chief technology officer and co-founder at Secure-IC, a Cadence company. “The notion of how to ensure trust end-to-end across the supply chain, across an ecosystem, has had the concrete effect of requiring ‘passibility.’”
Do not clone this
Following nearly two decades of slow growth, physically unclonable functions (PUFs) have emerged as a cornerstone technology for supply chain tracking. “PUF was leveraged so that you have an anchor for supply chain traceability and identity,” Guilley said. “As the system is getting manufactured, tested, and assembled, you end up with something off the shelf that you can sell and mount on racks where you have something a bit like a passport. You have the full history of the identity and how you pass ownership. This has been modeled in chain-of-custody technology. You can trace the provenance and identity, and all the actions performed on the hardware and software system, until you reach the final state where you can deploy and kickstart your data center. This happens when there is a real need for trustworthy systems that are built in an ecosystem that could have gaps or attackers between each manufacturing stage.”
That’s part of the solution. Data center defense systems also require continuous monitoring during runtime. “Hardware silicon security will continue to be a very important topic, but you have to make it anti-tamper,” said Dana Neustadter, senior director of product management for security solutions at Synopsys. “You must make sure that the identity is unique and cannot be compromised or cloned. That’s fundamental to building security solutions. Then you have firmware and infrastructure. You want to scan for vulnerabilities. Typically, here you use measured boot or real-time integrity monitoring. Today this is one of the areas that is the least monitored for attacks during runtime.”
More layers or better ones?
Multi-layered data center defense as an afterthought is officially a thing of the past. “Experts in security have always known this, and asked for it,” said Reed Hinkel, director of strategic programs at Synopsys. “It started more at the facility level, and then the system level. Then it was pushed down. From a silicon perspective, it was always treated as a best practice, but what does that mean? If it’s a best practice, that means it’s not a requirement; therefore, I don’t have to include it, or I don’t have to spend die area or software time or any of that.”
The industry-wide understanding is that data center security systems must be built with multiple, mutually reinforcing layers that leave no gaps across the many systems that govern a data center, a concept called defense-in-depth. “You need to have this trust around every single component and components within components, like confidential compute, hardware silicon security for firmware and infrastructure, network security, access management control, virtualization, AI model protection, and training infrastructure protection,” said Neustadter. “Everything needs to be protected, and whatever connection and dialog those components have needs to be authenticated and proven that, ‘Yes, I can communicate. I can share this data.’ That’s a fundamental shift in data center security now.”
Adding more layers is only part of the solution. The existing layers need to be reinforced, as well. “Traditional defense-in-depth strategies focused on adding more layers of protection,” said Ajay Kapoor, director of product management, Silicon IP at Rambus. “Today, the emphasis is shifting toward making those layers intelligent, adaptive, and continuously verifiable. Security architectures are increasingly built around zero-trust principles and an assume-breach mindset, where the goal is not only to prevent attacks, but also to limit their impact when they occur. This is further complemented by extensive telemetry and continuous monitoring and requires security architectures that can evolve over time. For this, hardware roots of trust, cryptographic agility, and secure update mechanisms are critical because they allow systems to respond to emerging threats and new cryptographic requirements without requiring complete redesign.”
Expanding boundaries of trust
Securing today’s data center supply chain ultimately hinges on expanding the boundaries of trust wider than ever before.
“How do you know that what you’re installing is really what you think it is?” Keysight’s Register said. “If you’re building a data center or you’re a hyperscaler, there’s no way you can individually test everything that comes into your data center. How do you think about the provenance of the actual hardware that’s coming in? How do you guard against counterfeits in the supply chain?”
Expanding the boundaries of trust means greater complexity, which creates some downstream costs. “The security has multiple costs,” Cadence’s Guilley said. “First, there is the hardware footprint. A root of trust in the past was very simple, with basic isolation. Now, cryptography is getting fat. We have post-quantum cryptography, and crypto-agility to support multiple algorithms, key sizes, and modes of operation. In the past, it was audited a little bit casually, but now it must be certified. This incurs a cost in verification, but also a cost in silicon, of course, and a cost to deploy, properly configure, and subsequently verify.”
Expanding the boundaries of trust also includes securing the data center’s AI models themselves. Infineon’s Laackmann pointed to several essential components:
- Defending the data input from poisoning;
- Protecting the model from tampering;
- Defending against manipulation of the data processing step; and
- Verifying the AI output.
This process can require multiple layers of digital signing. “The model itself can be digitally signed by using a secure element,” he said. “Thereby unwanted manipulation can be detected, and attacks can be prevented. Security-hardened MCUs allow protected monitoring and verification of the ongoing AI computing process. Hardware-based security features protect against logical attacks that target model drift or anomaly detection. To provide appropriate indication that a data set really originated from a specific AI or from a dedicated processing unit, the output data can be digitally signed, as well.”
New standards driving progress — and bottlenecks?
Open-source collaborations like Caliptra, the Open Compute Project Foundation’s root-of-trust project, are helping build hardware-level trust. “We now have some early systems like Caliptra to establish lower-level hardware roots of trust, which can be super important as we move on from old-school system-on-chip,” Register said. “I’m now using composable hardware with chiplets that need to talk to each other. Little pieces of those different chiplets may come from different manufacturers, but they must trust each other, and you have to have trust in each of those things that’s been composing a piece of hardware.”
One reason for Caliptra’s widespread adoption is the need to comply with NIST’s standards for data centers that handle sensitive government workloads, including looming PQC transition deadlines. “The primary drivers behind the secure boot and the remediation capability are certain NIST standards that need to be complied with by data centers to be operated as government cloud workload carriers,” said Hinkel. “Another aspect of why Caliptra came about was the executive order on the timeline for launching post-quantum cryptography. Data centers are on an accelerated path where anything that runs GovCloud, beginning next year, has to be post-quantum secure.”
PQC adaptability is especially important in the data center security space because many components deployed today will still be in use when hard PQC deadlines arrive. “Data center operators recognize that systems deployed today may still be responsible for protecting sensitive information a decade or more into the future, making cryptographic agility a critical design objective,” said Rambus’ Kapoor. “Therefore, organizations are increasingly seeking architectures that can accommodate future cryptographic transitions rather than hard-coding support for specific algorithms. Hardware and firmware must support algorithm updates and evolving standards without requiring wholesale system replacement. As PQC adoption accelerates, the focus shifts from simply selecting algorithms to ensuring infrastructure can evolve as standards and threats continue to develop.”
But while the need to certify products to meet government standards has driven some innovation, a backlog in product certifications creates its own supply chain problems. “The big issue is that now there are a lot of stragglers, and you’re basically forced to make sure that when you bring your post-quantum solutions to market, all are FIPS 140-3 compliant. The big challenge is that they’re already running a huge backlog,” Hinkel said. “It’s only going to get worse because everyone’s put off getting things certified, and that’s creating an immense problem. For anyone who waits any longer, things beyond your control will prevent you from getting your product out on time. It’s going to be sitting on a shelf because the government can’t buy it.”
Where physical meets digital
Experts see some of the most troubling attack surfaces at the meeting point of physical and cyber security. Potential threats to HVAC, liquid cooling, and energy systems require significant collaboration between IT and OT security professionals.
“If you just think of it on the national level, we are not taking the threat of cyber-physical seriously enough,” Register said. “There’s a fundamental lack of understanding of how dependent everything is, from our AI data centers and emergency management systems to health care systems and things like that. They are on fairly brittle Critical National Infrastructure systems. This is true whether you’re looking at that on the individual data center level or the broader national level.”
Concerns are also growing about cyber-physical attack chains. “Increasingly, operators are concerned less with individual attacks and more with attack chains that exploit multiple weaknesses across physical, cyber, and operational domains,” Kapoor said. “Threats such as physical tampering, supply chain compromise, credential theft, firmware attacks, ransomware, and large-scale DDoS campaigns remain particularly concerning because they target foundational trust boundaries and can be difficult to detect early. As attackers become more sophisticated, protecting the integrity of systems from manufacturing through deployment and operation has become just as important as defending against network-based attacks.”
Timing-based attacks targeting data centers can also be among the most debilitating threats facing facilities. “One of the things that I’m really interested in right now is timing. A lot of elements in power, such as energy grids, are very dependent on everybody having the same clock, and they’re being synchronized, so that’s a sine wave,” Register said. “As SDRs, radios, get cheaper and drones get cheaper, it becomes much more feasible to think about large-scale attacks that either jam or spoof satellite signals. You don’t have to get someone’s clock off by a week, a day, a minute, an hour, or even a second. I’m talking milliseconds off, and these bad things will happen. How we build in resiliency to things like that doesn’t get nearly enough attention. That can be at the data center level or up to a more widespread level.”
These threats require cyber and physical security systems to be purposefully integrated to defend against this new wave of threats. “While remote physical attacks are on the rise, it is no surprise that digital, or logical, security systems must be significantly merged with physical security in terms of detection, response and, last but not least, prevention,” Laackmann said. “For example, sensors and cameras can be equipped with secure elements that provide trusted authentication of the devices and the data they provide. Security-hardened MCUs allow secure processing, encryption, storage, and distribution of sensitive data, thereby drastically reducing the efforts needed to secure these physical protection systems.”
Related Articles
Keeping Security Algorithms Current Is Getting Harder
As threats evolve faster, protecting security algorithms from design through manufacturing and across the supply chain is becoming more difficult.
Data Center Security Issues Widen
The number and breadth of hardware targets is increasing, but older attack vectors are not going away. Hackers are becoming more sophisticated, and they have a big advantage.
Data Center Chokepoints Tied To AI, Political Pressure, Supply Chain
Technical, political, and supply chain challenges are complicating data center buildouts around the U.S.
The post AI Is Forcing Data Centers To Rethink Trust appeared first on Semiconductor Engineering.