This content was last updated in September 2024, and represents the status quo as of the time it was written. Google's security policies and systems may change going forward, as we continually improve protection for our customers.
Titanium hardware security architecture serves as the foundation for Google's services and underpins many of the security countermeasures in Google infrastructure. Titanium hardware includes security microcontrollers, hardware adaptors, and offload processors that were developed specifically to address specific attack vectors for Google infrastructure.
Titanium hardware is the latest advancement to Google's ever-evolving and comprehensive infrastructure security and helps protect the integrity, confidentiality, and availability of user data. Titanium hardware builds on infrastructure such as cryptographic hardware offload cards that provide encryption-in-transit and internal microservices that provide data-at-rest encryption.
This document describes how Titanium hardware components work together to create a security architecture that helps protect the physical attack surface of Google systems and mitigate threats to customer data. This document also describes how Titanium hardware enables specific security controls at the software layer, the evolution of the architecture beyond initial cryptographic hardware offloads, and the real-world threats that the Titanium hardware security architecture is designed to mitigate across Google's customer and deployment base.
Titanium hardware security architecture
The Titanium hardware security architecture is designed to defend against a spectrum of scenarios and threat actors. The following architecture diagram shows Titanium's independent, yet interlocking, components.
The Titanium hardware security architecture includes the following components:
- Caliptra root of trust for measurement (RTM): helps enforce a security perimeter for each silicon package. Caliptra RTM provides attestation and a unique ID to root cryptographic services.
- Titan chip RoT: interposes between a platform's boot flash and its main boot devices such as the baseboard management controller (BMC), the platform controller hub (PCH), and the CPU. Titan chips provide a physically tamper-resistant hardware-based RoT that helps establish a strong identity. Titan chips also help with code authorization and revocation for machines, cards, or peripherals.
- Titanium offload processor (TOPS): provides cryptographic controls to help protect the confidentiality and integrity of data at rest and data in transit.
- Custom motherboards: provide resiliency at scale against DoS attacks from faulty or malicious software, as well as protection against physical attacks. In the diagram, for example, the chip package and Titan RoT are on a custom motherboard that is separate from the custom motherboards for Titanium TOP or the motherboards for other infrastructure.
- Confidential Computing enclaves: help enforce isolation against Google's administrative privileges, improve isolation with other tenants, and add verifiability using attestation. Attestation can give assurance that the environment has not been altered.
- Backend fault-tolerant regionalized services: help prevent escalation of privilege across services, zones, or from administrative access.
In the diagram, Other infrastructure refers to network fabrics and replicated backend storage.
Design principles of Titanium hardware security architecture
Our Titanium hardware components and their interactions are developed using the following fundamental principles:
No single point of failure: Google architecture is designed to avoid single points of failure, such as single load-bearing components with multiple responsibilities. Building Secure and Reliable Systems discusses the importance of avoiding single points of failure. This principle is applied throughout our physical infrastructure, across all regions, and even down to the silicon in chips. This resiliency throughout our global infrastructure helps keep customer data safe and available.
For example, live migration with Confidential Computing helps to preserve the encrypted memory on supported host machines. Live migration helps to ensure that a long-running VM is not a single point of failure due to maintenance events or responding to vulnerabilities.
The perimeter is the silicon package: because a server system contains multiple interconnected and separate system-on-chips, our architecture fundamentally distrusts all connectors, fabrics, printed circuit board assemblies (PCBAs), PCBA traces, and cables. Though component separation is useful for modularity, separation can also become a weakness when it offers adversaries well-defined targets from which to snoop plaintext data. Data within the silicon package itself is encrypted and authenticated by private cryptographic assets within the package.
Moving the perimeter into the silicon itself helps minimize implicit trust. This principle addresses threats against data confidentiality that occur as data center serving conditions become increasingly diverse. For example, setting the perimeter into the silicon package helps address threats from rogue hardware operations.
Zero trust and compartmentalizing risk: multi-party controls on administrative actions help ensure that no personnel account (or compromised personnel account) can unilaterally cause the threats that are discussed in this paper. The architecture separates services into layers and zones to compartmentalize and contain risk. Even with enclaves, which are typically hardware rooted, the architecture accounts for the discovery of hardware vulnerabilities and the need to remediate while components remain in operation.
For example, if an attacker maliciously compromises a chip's behavior inside an active system that is running customer workloads in our data centers, the architecture is designed to identify and isolate that compromised chip. That machine can then be taken out of service. Even if the machine isn't taken out of service, the attacker must defeat additional trust boundaries and compromise multiple credentials to move laterally and gain influence over further systems, users, or regions.
Independent failure domains and isolation technologies help limit the affected area of a compromise. These domains and technologies add natural control points for detection, and limit the amount of additional complexity that must be introduced.
For more information about our zero-trust implementation, see BeyondProd.
Transparent security: Google invests in multiple transparency efforts, including open sourcing, responsible disclosure of research and findings, and partnering with the ecosystem of hardware manufacturers. Google global infrastructure applies Kerckhoffs' principle, which states that a cryptosystem should be secure, even if everything about the system, except the key, is public knowledge.
For example, we contribute to open source projects and use these projects in our security hardware designs and our security software. The following table describes some of the open source projects that we contribute to and use.
Open source project Description Titanium component Used in FIPS 140-3 level 1 encryption libraries
BoringSSL
Used in roots of trust (RoT) at the silicon level
Caliptra RTM
Used in RoT for a chip in a system architecture
Titan chips
Used for coverage-guided kernel fuzzing
Host ring0 and User VM Distros
Used in FIPS 140-3 level 1 encryption libraries
Titanium Offload Processor
Physical and logical defense in depth: Google relies on physical data center security to help protect our capital investments and our systems. These security controls are an initial layer of defense, so we deliberately invest in additional logical controls to harden our systems against physical threats. Titanium adds to our defense in depth by adding compartmentalization in our hardware that provides additional defenses against specific infrastructure threats.
For example, our data centers have metal detectors that can accurately detect storage media exfiltration attempts. However, our data-at-rest encryption strategy is deliberately engineered not to rely on physical media custody. These logical and physical controls are independent and complementary layers.
Our combined physical and logical security controls help us stay vigilant against insider threats and help protect the confidentiality of our users' data.
Security benefits of Titanium architectural components
The following table highlights some important security benefits that are achieved with the Titanium security architecture components, both at the hardware and software layer. These security benefits are described in more detail in the following sections.
| Security benefits | Architecture component |
|---|---|
Silicon-level trust perimeter on system-on-chips (SoCs) such as CPUs or GPUs |
Caliptra RTM |
Silicon-level verifiability |
Caliptra RTM |
Cryptographic identity at the hardware level |
Caliptra RTM, Titan RoT |
Verification that the expected binaries are running |
Caliptra RTM, Titan RoT |
Persistent threats mitigation across boots |
Caliptra RTM, Titan RoT |
Protection of the confidentiality of data at rest and data in transit |
TOPs for cryptography |
Offloading processor-level protection (beyond a physical card) |
TOPs for cryptography |
Security by design, resistance to physical attacks, and resilience capabilities supporting full system firmware recovery from a single Titan RoT |
Custom motherboards |
Purpose-built boards with only essential connectors, which helps mitigate against physical tampering attempts |
Custom motherboards |
Cryptographic workload isolation from machine-wide system software and administrative human access |
Confidential Computing enclaves |
Tamper resistance through DRAM encryption (to enable data-in-use encryption) |
Confidential Computing enclaves |
Minimized affected area and bulkhead for an attacker with local access |
Backend fault-tolerant regionalized services |
Multi-levels of compartmentalization |
Backend fault-tolerant regionalized services |
Caliptra root of trust for measurement
Caliptra RTM helps build trust and transparency for the firmware within our ecosystem that runs on system-on-chips (SoCs) such as CPUs, GPUs, and TPUs.
Caliptra RTM has the following benefits:
- Provides a root cryptographic service: Caliptra RTM helps detect corrupted critical code and configuration through strong end-to-end cryptographic integrity verification. Caliptra RTM can cryptographically measure its code and configuration, sign these measurements with a unique and hardware-protected attestation key, and report measurements that attest to the authenticity and integrity of the device. Caliptra RTM provides a cryptographic device identity and a set of firmware and configuration integrity measurements for the motherboard.
- Mitigates physical supply chain security: Caliptra RTM helps ensure that the hardware is authentic and is running intended firmware and software. In combination with software supply chain security, Caliptra RTM lets the system verify the authenticity and integrity of firmware and software, whether created by Google or a third-party. This verification process lets Caliptra RTM maintain authenticity and integrity across authorized updates and helps ensure that configurations remain as intended and are attested.
- Protects against physical intrusions that require direct access to running hardware: Because Caliptra RTM is built into the chip's silicon layers, a PCBA interposer or rogue chip that tries to deliver the wrong firmware to an application-specific integrated circuit (ASIC) cannot successfully attack the RoT. For example, attackers can bypass an external RoT's detection capabilities by tampering with the relatively low-speed SPI bus. However, an RoT that is embedded within an SoC or ASIC becomes more difficult for an attacker to compromise.
Titan chip root of trust
Titan is designed to cryptographically maintain device identity, defend against bad software pushes, and enforce code authenticity using revocation.
A strong cryptographic device identity helps ensure that the fleet is composed exclusively of validated machines that are running the expected binaries and can identify and authenticate legitimate access. Legitimate access is rooted at the hardware level.
By default, production machines use trusted boot to help ensure that only authenticated software can run. Trusted boot verifies the digital signature of all boot components and doesn't permit a machine to participate in the production environment if authentication fails.
As an additional preventative control, machine code revocation prevents software changes that are no longer authorized from being applied. The revocation capability in Titan chips helps mitigate not only malicious attacks (for example, rollback or replay attacks), but also non-malicious stability or resilience bugs (for example, by preventing accidental reinstallation of buggy old firmware).
For more information, see How Google enforces boot integrity on production machines.
Titanium offload processors for cryptography
Titanium offload processors (TOPs) for cryptography help provide security during offloading of I/O. These TOPs are protected with Titan or Caliptra RTM. TOPs deploy widespread authenticated encryption of data in transit and authenticated encryption of data at rest at low cost. Authenticated encryption means that customer data has cryptographic assurance of confidentiality and integrity. Because TOPs manage the cryptography, they deprivilege many system components. TOPs enable enhanced architectural properties, like availability, while minimizing the potential for loss of data secrecy.
Custom motherboards
The custom motherboards in Google infrastructure are designed to provide hardware provenance. The motherboards support attestation at multiple layers. Motherboard designs safeguard customer data, even in the highly unlikely event that an attacker physically attaches a malicious device to a machine. Titanium motherboard designs help enable reliable deployment of additional hardening mechanisms, such as depopulated debug ports, read-only serial consoles, bus connector intrusion, and extrusion signaling.
TLS and ALTS are the only accepted protocols exposed by our BMC network stack when a machine is turned on. For machines that use a third-party COTS design, such as our X4 instances, TOPs proxy any management traffic that is unique to that third-party design. Proxying management traffic means that our infrastructure doesn't depend on third-party designs for authentication, authorization, transport security, or network security.
Titanium custom motherboards are designed to have built-in recovery and backup mechanisms to ensure availability and recoverability. They can restore themselves from most crashes or from firmware corruption. Our latest designs enable rebuilding the entire machine from a single working Titan RoT. These motherboards use feature-dedicated power componentry and reset signaling to help ensure the electrical independence of Titan RoTs from the rest of the platform, and help protect their control over the platform's firmware payloads for the purposes of authentication and recovery.
Confidential Computing enclaves
Confidential Computing creates a Trusted Execution Environment (TEE) or enclave to help isolate customer sensitive workloads from Google administrative access. When the data is handled by the CPU or GPU, Confidential Computing provides a technical preventative control through compute isolation and in-memory encryption. Confidential Computing helps ensure that even a malicious hypervisor cannot access a VM. For customer workloads, Confidential Computing provides a layer of data secrecy insulation from the possibility of unintended Google personnel access or faulty automated system-software actions at scale.
One example of advanced security that is enabled by the Titanium architecture is Confidential mode for Hyperdisk Balanced. Confidential mode for Hyperdisk Balanced combines Titanium-based block storage offloads, Confidential Computing, and Cloud HSM to create a hardware-rooted TEE. In other words, Confidential mode for Hyperdisk Balanced is a hyperdisk-balanced offering. Confidential mode for Hyperdisk Balance isolates infrastructure so that sensitive keys are exclusively processed in a hardware-rooted TEE. For information about the third-party review of the cryptographic operations, see Public Report – Confidential Mode for Hyperdisk – DEK Protection Analysis.
Backend fault-tolerant regionalized services
Backend fault-tolerant regionalized services help minimize the affected area from an attacker with local access. Google infrastructure is designed to compartmentalize services, systems, and zones from lateral movement of privileged insiders or compromised services.
We are working to include regional information in an increasingly broad set of our internal identity and access management systems. Regional information strengthens cryptographic isolation so that an attacker who obtains local access must compromise multiple credentials from distinct infrastructure services to continue to move laterally.
If an attack triggers a preventative control that would take a production machine out of the environment (for example, cause the system to power off), our fault-tolerant backend infrastructure helps ensure the continued availability of customer data and services on nearby machines. For more information about our infrastructure controls, see BeyondProd and How Google protects its production services.
Attack vectors for Google Cloud infrastructure
This section describes specific physical and logical threats that make up some of the attack surface of Google Cloud. The Titanium hardware security architecture is specifically designed to address a unique set of threats to Google infrastructure and the user data that we store.
Infrastructure threats
Titanium architecture is designed to defend against the following multiple categories of threats:
- Rogue insider with physical access: Our personnel need access to physical devices in data centers to deploy, maintain, and repair hardware. This access represents a potential attack vector because rogue personnel or contractors have a legitimate business reason to physically repair some of the machines in our data centers.
Rogue insider with logical access: Similar to physical access to the data center, personnel are required to develop, maintain, test, debug, tune, and support multiple levels of the Google software stack. These personnel include developers, SREs, and customer-facing cloud engineers.
For more information about our defenses against this threat, see How Google protects its production services.
External attacker with logical access: External attackers can gain a foothold inside of a Google Cloud environment and attempt to transfer laterally to other machines to obtain access to sensitive data. A common tactic used by external attackers is to begin by compromising a legitimate personnel or contractor account.
The following diagram shows which part of the cloud environment is most vulnerable to these threats.
Attack surface to data center servers
The following table describes attack surfaces that are typical aspects of data center servers. The Titanium hardware security architecture is designed to provide strong defenses against such threats.
| Attacker | Target | Attack surface | Risk |
|---|---|---|---|
Rogue insider with physical access |
Storage media (SSDs, HDDs, or boot drives) |
Physical drives and connectors |
This attack could steal a drive, and attempt to access it with the attacker's tools. |
DIMM |
Physical memory connectors |
This attack could freeze the DIMM, take it out of the data center, and try to access the data on it using the attacker's own tools. This threat is sometimes called a cold boot attack. |
|
Server |
USB or PCIe connectors |
This attack could connect malicious hardware to the server. Using the malicious hardware, the attacker could attempt to gain code execution or exfiltrate resident data. |
|
Motherboard |
Joint test access group (JTAG) eXtended Debug Port (XDP) |
This attack could connect a hardware debugging tool to gain code execution or access to data that is processed on the CPU. |
|
Network |
Ethernet cables |
This attack could tap an ethernet cable to gain access to all the data that is being transferred between devices. Any cleartext traffic could then be observed. |
|
Motherboard |
Firmware |
This attack could introduce persistent malicious firmware. This firmware could be pre-installed from a compromised manufacturer, interdicted in transit, or updated by an insider. This threat can lead to pre-hacked hardware with rootkits that provide backdoor access to the server. |
|
Rogue insider with logical access |
Compute workload (for example, VMs) |
Login points |
This attack could use insider credentials to directly access VMs or hosts, and the data on them. |
Fabric router |
Physical or administrative access |
This attack could gain root control over a fabric router to eavesdrop on all traffic and exfiltrate or tamper with any cleartext data that is in transit on the fabric. |
|
Motherboard |
Firmware |
This attack could push defective firmware images to motherboards, making them permanently inoperable and rendering data unrecoverable. An attacker could push known-vulnerable firmware to machines to regain control using exploits that enable remote code execution. |
|
External attacker with logical access |
Server |
VMs |
This attack could launch public side-channel attack patterns on VMs. These attacks could leak data from instances running on the same hardware, or from host system software. |
SSDs |
VMs |
This attack could use direct access to PCIe SSDs to attempt to infer co-tenant data. |
|
Memory |
VMs |
This attack vector could use side channels to search memory for valuable encryption keys. |
|
Server |
Bare-metal VMs |
This attack vector could use bare-metal instances to scan all peripherals to find a vulnerable component that would let them persist in the machine and attack subsequent tenants. |
Mapping Titanium hardware components to threats
The Titanium hardware security architecture uses a multi-layer approach to help address specific infrastructure threats and to prevent single points of failure. These threats may originate from mistakes or from rogue actors. The threats span hardware operations and can exploit vulnerabilities in servers, networks, and the control plane. No single solution exists that can address all of these attack vectors, but the combined features of Titanium help protect our users data and our cloud computing instances.
Scenario: Rogue hardware operations
Rogue hardware operations pose a threat to data security, as they can lead to the exfiltration of data from data centers and the modification of hardware and firmware. Google's Titanium hardware security architecture helps defend against these threats by using a variety of security measures, including cryptographic RoTs, custom motherboards, and I/O processors. These components work together to provide a layered defense that is resistant to a wide range of attacks.
The following table describes some of the rogue hardware threats and how Titanium architecture can mitigate them.
| Threat | Titanium mitigation |
|---|---|
An attacker exfiltrates individual data drives from the data centers to access the data on them. |
Data-at-rest encryption keys for storage products and services are never stored persistently on the machines to which storage media is attached. The built-in self-encryption capabilities of storage media are also enabled for defense in depth, and use keys that are never stored persistently on the media itself. Caliptra RTMs let Google include root-of-trust hardware identity and firmware integrity among the authorization conditions that are required to release keys from a key management service to storage service instances. Machines that are somehow maliciously configured with unintended firmware can't access the keys that are needed to decrypt stored data. The RoTs embedded inside the silicon packages anchor the relevant cryptographic identities within the chip package. Single-function interposers are the main part of our data plane security and encrypt data at every processing step. TOPs provide the following benefits:
Proven software solutions like dm-crypt are used for lower-performance drives where reducing the attack surface is paramount, such as some boot drive use cases. |
An attacker taps a network cable and reads bytes on the wire or fiber. |
TOPs encrypt data in transit, which removes the opportunity for a threat to sniff valuable data on the network. Our NICs use the PSP hardware offload standard. This standard provides cost-effective encryption with minimal decrease in performance. These implementations are FIPS compliant. Customer data is encrypted when it transits Top of Rack (ToR) or fabric switches. Some machine-learning infrastructure uses proprietary transport security mechanisms. |
An attacker replaces flash chips that hold mutable code in the data center or the supply chain to run malicious code on the servers. |
Titan chips are designed to reject the attack and don't provide access to the credentials that are stored inside. Even if an attacker rewrites the content of non-volatile flash chips, the Titan RoT securely reports a measurement of the code to Google's control plane, which is designed to block the device. Google routinely revokes deprecated or known-vulnerable code at global scale in our fleet by using Titan chips. |
An attacker inserts adversarial devices in physical interfaces of data center servers or cards to run malicious code or exfiltrate data. |
Custom motherboard designs remove the interfaces that are used to insert adversarial devices. Input-Output Memory Management Unit (IOMMU) configurations are in place to prevent PCIe screamers in all of our firmware. (PCIe screamers are designed to read and write arbitrary packets on the PCIe fabric.) As the industry matures, we are complementing this protection with PCI IDE to additionally mitigate against more sophisticated PCI interposers. ALTS and TLS are the only accepted authentication and authorization network connections for control and management functions on TOPs and BMCs. Caliptra RTMs block any non-approved firmware. Our trusted peripherals attest their hardware identity and code integrity to our control plane, and no server is admitted to production if the attestation record does not match the hardware and software intent. |
An attacker uses a cold boot attack in the data center to access data in RAM. |
Confidential Computing in-memory encryption protects any sensitive data or encryption keys in RAM. DRAM encryption is also enabled in machines that are deployed without Confidential Computing in lower assurance edge data centers. |
Scenario: Exploitation of servers or networks by rogue users
Attackers might use the public cloud to host their malicious workloads on our shared infrastructure and to deposit data in our public services. External adversaries, from single individuals to nation states, might also attempt to gain remote privileged access.
To mitigate these actions, the Titanium hardware security architecture uses Titan chips and Caliptra RTM to provision runtime credentials in a secure manner and limit privileges on hardware and operating systems. Confidential Computing helps protect against manipulation of system memory – whether physically or using hypervisor attacks – and Titan chips reject or detect unauthorized software upgrades.
The following table describes some of the server and network exploitation threats and how Titanium architecture can mitigate them.
| Threat | Titanium mitigation |
|---|---|
An attacker exploits a vulnerability to escape their VM and gains access to data and other VMs running on the same machine. |
Confidential Computing enclaves curtail the exfiltration of workload data, whether in process or at rest. This mitigation blocks a VM-escaped attacker from accessing data in use. Titan chips and Caliptra RTMs block the attacker from having persistent access. Any attempts for persistent access are likely to be detected because the machine configuration won't match the configuration and code policy for that server. This match is required before the machine can host production workloads after a reboot. |
An attacker launches public side-channel attack patterns on VMs. |
Our fleet management system, using Titan chips, can revoke known-vulnerable software. Revocation can block any subsequent attacks that target these known vulnerabilities. Titan-based integrity measurements also provide high confidence that mitigations, which might need to be urgently deployed, have been deployed to the target machines. We reinforce this approach by remaining at the forefront of side-channel investigation and mitigation, through techniques like retpoline and core scheduling, and advanced research on Meltdown, Spectre, Zenbleed, Downfall, and others. |
An attacker uses direct access to SSDs that provide storage to multiple tenants to attempt to infer co-tenant data. |
Data-at-rest encryption helps protect against logical and physical attacks with a variety of interposers. For resources that aren't shared, each tenants' data is encrypted using different keys, which mitigates the opportunity for direct access attacks against the SSD. |
An attacker scans memory and uses side channels to search for data encryption keys or credentials. |
Titan chips enable provisioning of per-machine sealed credentials. Even if an attacker gains root access on one machine, its credentials are bound solely to the private identity of the local Titan chip. |
An attacker purchases bare-metal instances and scans all peripherals to try to gain persistent access. |
Titan chips reject any unauthorized software upgrade, including malicious pushes for persistent control. Our machine workflow positively confirms expected integrity measurements across a full system attested power cycle between bare-metal customers. |
Scenario: Exploitation of servers or networks by rogue control plane behavior
Rogue control plane insiders can attempt to exploit Google's systems in a number of ways, including attempting to gain root control over a fabric router, pushing defective firmware images to motherboards, and eavesdropping on network traffic. The Titanium hardware security architecture defends against these threats by using a variety of mechanisms, including Titan chips, Caliptra RTMs, custom motherboards, and backend fault-tolerant isolated services.
The following table describes some of the control plane threats and how Titanium architecture can mitigate them.
| Threat | Titanium mitigation |
|---|---|
An attacker uses insider credentials to access Compute Engine VMs that are serving as the foundational layer for customer environments. |
TOPs help ensure that administrators don't have access to customer environments. Without access, Google personnel can't use their credentials to access the privileged hardware and software layer that is underneath our customers' VMs. Google insider access to customer data is blocked because data is only accessible through defined APIs. |
An attacker pushes defective firmware images at scale to motherboards, permanently bricking them. |
Titan chips RoTs reject any unauthorized software upgrade, including malicious pushes for persistent control. Custom motherboard designs use an alternate network of signals that interconnect all of our RoTs to the platform RoT. The platform RoT holds backup firmware for critical devices. Even if networking and PCI were bricked by an attacker, the out-of-band (OOB) network can heal the system. |
An attacker pushes deprecated known-vulnerable production firmware to machines to regain control using public vulnerabilities. |
Titan chips reject bad pushes and help enforce revocation of known-vulnerable code. They attest the firmware version that is deployed on the machine and reject the machine at the control plane. This mitigation helps prevent any jobs from running on an unhealthy machine, and triggers investigation or repair as necessary. |
An attacker abuses silicon debugging capabilities that are necessary for business continuity, which provide the highest conceivable level of data access in server systems. |
Caliptra RTM helps ensure that all parameters that enable invasive debug interfaces, whether logically connected or through direct physical insertion, are trustworthily configured, cryptographically measured, and reported to our control plane using an attestation protocol. Only machines in the intended state gain access to serve production workloads. |
An attacker gains control of one backend service so that they can access customer environments. |
Backend fault-tolerant regionalized services are regionalized credential infrastructure that don't permit unilateral human access. In addition to preventing operator login onto compute nodes, operators also cannot log in to the control plane to retrieve key material. Confidential Computing enclaves in the Titanium architecture isolate our backend authorization and key provisioning services from machine root privileges. Key hierarchies help protect the signing and authorization keys of most services. With key hierarchies, the root keys are in air-gapped keys that are held in HSMs and in safes, or keys that are held in production by a Paxos quorum of in-memory data stores. |
What's next
- Read the Infrastructure security design overview.
- Implement Confidential Computing.