Prepare and test your skills
Prepare and test your skills
Secure-by-design infrastructure means Google builds security into the hardware, software, network, and data centers that run Google Cloud (opens in a new tab) instead of adding it later. The shared responsibility model divides security duties between Google and the customer, and the exact split depends on the service model. Together they explain what Google protects for you and what you must still protect in an AI environment.
The shared responsibility model divides cloud security into protecting the cloud itself and protecting what the customer puts in the cloud. Google is always responsible for the underlying infrastructure, while the customer is always responsible for their own data. Google Cloud provides security services that help customers secure their environment at scale, but using those services does not remove the customer's own obligations.
Where the line falls changes with the service model. In infrastructure as a service (IaaS), Google secures only hardware, storage, and network, and the customer secures everything above those layers. In software as a service (SaaS), Google secures everything except the data and its access and usage.
| Service model | Google's responsibility | Customer's responsibility |
|---|---|---|
| IaaS | Hardware, storage, and network | Everything above hardware, storage, and network, including data |
| SaaS | Everything except data and its access and usage | Data and its access and usage |
Exam tip: In IaaS, Google secures only hardware, storage, and network; in SaaS, Google secures everything except your data and its access and usage.
Secure-by-design infrastructure is Google's practice of making security a primary design rule for its own servers, operating system, network, and data centers. Google follows defense in depth, which layers multiple controls so no single failure exposes data. Because Google runs Search, YouTube, and Gmail on the same infrastructure it sells to customers, customers inherit the same controls and practices.
Physical security at Google data centers uses layered safeguards that get stronger closer to the data center floor. Outer layers include perimeter fencing, vehicle access barriers, alarms, metal detectors, custom electronic access cards, and biometrics. The buildings use laser beam intrusion detection, 24/7 monitoring by high-resolution interior and exterior cameras, and guards who have passed background checks and training. Entry to the data center floor requires passing through a security corridor with multi-factor access control using badges and biometrics, and only approved employees with specific roles may enter.
Inside the data center, Google applies controls in the physical-to-logical space, which is the distance from a machine in a rack to that machine's runtime environment. These controls include hardware hardening, task-based access control, anomalous event detection, and system self-defense. Google's servers are purpose-built and omit unnecessary parts such as video cards, chipsets, and peripheral connectors that could create vulnerabilities. Custom chips such as Titan help Google identify and authenticate legitimate Google devices at the hardware level, including the code those devices use to boot.
Hardware tracking and disposal follow strict rules to protect data at rest and at retirement. Equipment location and status are tracked with barcodes and asset tags, while metal detectors and video surveillance help prevent unauthorized removal from the floor. Storage devices such as hard drives, solid-state drives, and non-volatile dual inline memory modules (DIMMs) use full disk encryption (FDE) and drive locking. When a device is retired, authorized staff verify sanitization through multiple-step verification, and if data cannot be erased the device is physically destroyed with a shredder and recycled at a secure facility.
Software development follows the same secure-by-design approach from source code to deployment. Google uses source control protections and two-party reviews to reduce vulnerabilities, plus libraries that prevent whole classes of bugs such as cross-site scripting (XSS) in web apps. Google checks source code, build process, and code provenance with the Supply-chain Levels for Software Artifacts (SLSA) framework. Customers can enforce workload provenance with Binary Authorization (opens in a new tab) and use only verified open-source libraries with Assured Open Source (opens in a new tab).
Operational security is the set of ongoing practices Google uses to find, prevent, and respond to threats against its infrastructure. Vulnerability management, malware prevention, security monitoring, and incident management operate as continuous processes rather than one-time checks. Automation pipelines continually reassess patch deployment to fix vulnerabilities and flag incorrect or partial deployment.
Google's security monitoring program gathers signals from internal network traffic, employee actions on systems, and outside knowledge of vulnerabilities. A core principle is to aggregate and store security telemetry in one location for unified analysis. At many points across the global network, internal traffic is inspected for suspicious behavior such as botnet connections, and system logs are examined for unusual behavior such as attempts to access customer data. Automated network and log analysis escalates potential unknown threats to security staff, who also review inbound security reports and public mailing lists, blog posts, and wikis.
Incident management covers events that might affect confidentiality, integrity, or availability of systems or data. Google's program aligns with NIST guidance on handling incidents (NIST SP 800-61), and key staff are trained in forensics and evidence handling. Response plans for key areas, such as systems that store customer data, are tested against scenarios including insider threats and software vulnerabilities. The Google security team is available 24/7 to all employees, and if an incident affects customer data, Google or its partners inform the customer while the team investigates.
Control of access to customer data starts from a default rule: Google personnel do not have access to Google Cloud customer data. To obtain access, personnel must satisfy identity, device, role, and justification checks, and different roles receive different levels of access. Support roles have limited access tied directly to a support ticket, while engineering roles may need additional system privileges for complex reliability or deployment issues.
A Google employee must meet all of the following conditions before accessing customer data:
Google personnel may access customer data only in specific situations: customer-initiated support, outage or tool failure, third-party legal requests, and Google-initiated reviews. During an outage, engineers can access data without customer approval but cannot bypass the resource's allowlist, and the access is logged and audited. Third-party legal requests are rare, only the legal team can create a valid legal access justification, and Google notifies the customer when legally permitted. Google-initiated reviews are also rare and confirm that customer data is safe, such as removing malware that is exhausting a virtual machine's capacity.
Several Google Cloud services give customers visibility and control over Google's access. The following table compares what each service controls:
| Google Cloud service | What it controls |
|---|---|
| Access Approval | Requires customer approval before an authorized Google administrator can access data for support; approved requests are logged with Access Transparency logs linked to the approval |
| Access Transparency | Logs administrative access by authorized Google personnel for support or service availability |
| Assured Workloads | Provides workflow to create and monitor control packages for dedicated regional support, certified programs such as FedRAMP or ITAR, and Sovereign Controls for EU |
| Cloud KMS with Cloud EKM | Lets customer encrypt data with keys stored in a third-party key system outside Google's infrastructure, separating data at rest from keys |
| Confidential Computing | Encrypts data in use; includes Confidential VM, Confidential Google Kubernetes Engine Nodes, Confidential Dataflow, Confidential Managed Service for Apache Spark, and Confidential Space |
Compliance support is the set of independent audits, certifications, and attestations showing Google Cloud meets recognized security, privacy, and compliance standards. Google Cloud regularly undergoes independent verification of its security, privacy, and compliance controls. Key international standards include ISO/IEC 27001 for information security management, ISO/IEC 27017 for cloud security, ISO/IEC 27018 for cloud privacy, and ISO/IEC 27701 for privacy. SOC 2 and SOC 3 reports are available to customers, and Google participates in sector and country frameworks such as FedRAMP for the US government, BSI C5 for Germany, and MTCS for Singapore.
Certifications extend to AI models in a specific way: when a Google Cloud service includes Google models, the service's certifications also apply to those models unless explicitly documented otherwise. For regulated industries such as finance, government, healthcare, or education, Assured Workloads (opens in a new tab) helps implement industry-specific requirements inside Google Cloud. Google also maintains insurance covering many risk types, including cyber and privacy liability for unauthorized network use, failure to protect confidential information, notification costs, and crisis management such as forensic investigation.
Identity and Access Management (IAM) (opens in a new tab) principles are the rules that decide which identity can access which resource, for how long, and for what reason. In Google Cloud, privileged access is access by Google personnel to customer data to provide a contracted service, such as support or recovery from an outage. For generative AI, the same principles regulate access to sensitive training datasets, model endpoints, and administrative functions.
Least privilege means every identity receives only the permissions needed to do its job and nothing more. Google personnel are denied system access by default, and any granted access is temporary and no greater than necessary for the role. Personnel can reach only the resources needed for their work and must provide a valid justification to access customer data. This design limits insider risk, controls internal attacks, and reduces the impact of any incident.
Google enforces least privilege with infrastructure-wide controls that shrink attack paths:
Roles, permissions, and multi-party approval connect a job function to exactly what that identity is allowed to do. A role defines the level of access needed for a job function, and permissions attached to that role control what actions the identity can take. Support roles have limited access tied to a customer ticket, while engineering roles may have additional privileges for reliability or deployment. Access to customer data, critical operations on production systems, and changes to source code are controlled by manual and automated verification, so no one can act alone without approval from another qualified individual.
By default, Google personnel have no access to customer data. To obtain access they must meet strict conditions that combine identity, device, role, and justification:
Exam tip: Access Transparency logs Google personnel access after it happens, while Access Approval requires customer approval before that access is permitted.
Strong authentication verifies both users, such as employees, and roles, such as services, before they can reach data. Google requires strong passwords with security tokens or other two-factor authentication, and phishing-resistant sign-in with a Titan Security Key reduces the risk of stolen credentials. When a workload starts, it is provisioned with cryptographic credentials that it uses to prove its identity to other services with application layer transport security (ALTS). The ability to deploy or modify jobs that use a particular identity is restricted to those responsible for running the service, generally Site Reliability Engineers (SREs).
Context-aware access applies zero-trust checks so decisions do not rely only on static credentials or being inside a corporate network. The complete context of a request is evaluated, including user identity, location, device ownership and configuration, and fine-grained access policies. Machine inventory is securely maintained and each connecting device is evaluated for operating system updates, security patches, device certificates, installed software, virus scans, and encryption status. Authenticated users on trusted devices are then granted only limited, temporary access to necessary resources.
Logging, monitoring, and customer controls make privileged access visible and reviewable by both Google and the customer. Google personnel access to customer data is logged, and threat detection systems scan logs and match entries to threat indicators in real time. Strong authentication lets Google separate human events, service events, and service impersonation events to prioritize investigation of actual human access. Activities involving user data, source code, and sensitive information require a business justification or exception, with defined alerting for unilateral action on sensitive systems or access without a valid business reason.
Customers keep control through transparency and encryption tools that sit outside Google's internal permissions. The following table compares what each control records or protects:
| Control | What it records or protects | Customer control |
|---|---|---|
| Access Transparency | Logs administrative access by authorized Google personnel for support or availability | Customer can review logs and business justification reason codes |
| Access Approval | Covers requests by authorized Google administrators to access sensitive data | Customer must approve or deny before access is permitted |
| Customer-managed encryption keys (CMEK) | Protects data with keys managed by the customer in Cloud Key Management Service (Cloud KMS) (opens in a new tab) | Customer manages keys and controls access to them |
Security monitoring and threat detection services in Google Cloud collect signals from infrastructure, workloads, logs, and AI assets to help teams find, prioritize, and respond to attacks. Security Command Center (SCC) (opens in a new tab) is the central service that aggregates and prioritizes findings and recommendations. Workload monitoring tools feed visibility into SCC and Google Security Operations (Google SecOps) (opens in a new tab) for auditing and incident response.
Security Command Center (SCC) aggregates and prioritizes security findings and recommendations across Google Cloud environments. SCC gives security teams a central place to proactively address potential risks. SCC can integrate with ticketing systems, map users to Google SecOps, enable response playbooks, and ingest logs from Google Cloud, AWS, and Azure into Google SecOps. Security Command Center Enterprise (SCC-E) includes AI Protection, which provides a centralized view of AI assets, risks, and threats across the AI lifecycle.
Threat detection services such as Cloud Armor, Event Threat Detection, Container Threat Detection, and Virtual Machine Threat Detection scan logs and workloads and send findings to Security Command Center, which aggregates and prioritizes them. Prioritized findings flow into Google SecOps, where SIEM scans logs for threats and SOAR manages response workflows, with remediation results fed back as improved threat indicators.
AI Protection in SCC-E discovers AI inventory, including models, datasets, and applications, so teams know what needs protection. It uses virtual red teaming to identify attack paths and toxic combinations of misconfigurations and vulnerabilities. It also detects and responds to AI threats such as large language model (LLM) hijacking, data poisoning, and exfiltration, with response and remediation workflows that integrate with Google SecOps.
Threat detection services identify specific malicious activity and send findings to SCC for prioritization. Each service watches a different part of the environment, so coverage depends on using the right detector for each workload type:
Workload monitoring and visibility tools provide the raw activity data that security auditing and incident response depend on. Event Threat Detection monitors Cloud Logging and Google Workspace, while Container Threat Detection, Cloud Run Threat Detection, Virtual Machine Threat Detection, and Agent Platform Threat Detection watch runtime behavior in containers, VMs, and agents. Sensitive Actions Service watches organization, folder, and project actions for risky changes. Asset inventory from multi-cloud environments flows into the Google SecOps Entity Graph, which adds context to alerts.
Google Cloud's security monitoring program inspects internal network traffic for suspicious behavior such as botnet connections. System logs are examined for unusual behavior such as attempts to access customer data, and automated network and log analysis escalates unknown threats to security staff. Cloud Audit Logs track administrative activities and access to resources. Access Transparency logs administrative access by authorized Google personnel for support or availability, while Access Approval requires customer approval before an authorized administrator can access highly sensitive data, with approved requests logged and linked to Access Transparency logs.
Exam tip: Security Command Center aggregates and prioritizes findings, while Google SecOps SIEM and SOAR help teams detect, investigate, and respond to threats.
Security auditing and incident response use monitoring data to investigate policy violations and improve defenses. Google SecOps integrates with SCC to help detect, investigate, and respond to threats. Google SecOps includes security information and event management (SIEM), which scans logs and other data across clouds, defines detection rules, and searches accumulated data. It also includes security orchestration, automation, and response (SOAR), which manages cases, defines response workflows, and searches response data. Google SecOps uses AI and machine learning to analyze large volumes of security data, detect anomalies, and automate response, with dashboards showing time to detect and respond, DDoS mitigation effectiveness, and number of prevented attacks.
When policy violations are detected, security teams separate from core engineering conduct an initial investigation in a fixed order:
Remediation uses past incidents to resolve vulnerabilities and improve detection. All incidents are documented, and metadata is extracted to identify specific tactics, techniques, and procedures for each exploit. The team uses that data to develop new threat indicators, reinforce existing protections, or request security improvements. Incident management aligns with NIST SP 800-61, and the Google security team is available 24/7.
Access Transparency logs administrative access by authorized Google personnel after it happens for support or service availability, while Access Approval requires customer approval before an authorized administrator can access sensitive data. Approved Access Approval requests are logged with Access Transparency logs linked to the approval.
In infrastructure as a service (IaaS), Google is responsible for hardware, storage, and network and the customer is responsible for everything above those layers including data. In software as a service (SaaS), Google is responsible for everything except data and its access and usage and the customer is responsible for data and its access and usage.
Security Command Center (SCC) aggregates and prioritizes security findings and recommendations across Google Cloud environments, while Google Security Operations (Google SecOps) helps detect, investigate, and respond to threats. Google Security Operations (Google SecOps) includes security information and event management (SIEM), which scans logs and other data across clouds, and security orchestration, automation, and response (SOAR), which manages cases and response workflows.
Confidential Computing encrypts data in use in Google Cloud. Confidential Computing includes Confidential VM, Confidential Google Kubernetes Engine Nodes, Confidential Dataflow, Confidential Managed Service for Apache Spark, and Confidential Space.
An enterprise is planning to deploy a custom generative AI application on Google Cloud that processes proprietary corporate data. As part of their security architecture review, the leadership team must clarify how obligations are divided under the shared responsibility model.
Which security obligation is the direct responsibility of the customer when protecting their AI environment?
Last updated