Prepare and test your skills
Prepare and test your skills
Customizable Solutions and Agent Deployment is the use of pre-built and adjustable AI building blocks to tailor generative AI to an organization's needs while keeping architectural control. Cloud Run (opens in a new tab) is the serverless container platform that hosts these solutions, including intelligent agents. Because teams package their own code and dependencies in containers, they can connect the AI workload to other Google Cloud services and shape it for specific operational requirements.
Why Cloud Run fits AI workloads is its combination of flexible packaging, network access, scaling, hardware acceleration, and built-in connections to data and security services. A team deploys a Cloud Run service or instance and receives a secure Cloud Run URL endpoint that clients call over HTTP. Streaming for interactive AI responses is supported through HTTP chunked transfer encoding, HTTP/2, and WebSockets. The platform handles much of the operations work while the team keeps control of the container and architecture.
Cloud Run provides the following features for AI applications:
Exam tip: Cloud Run services scale automatically based on demand or can use manual scaling, while Cloud Run instances do not autoscale and run as singletons that are started and stopped manually.
Hosting AI agents and chatbots on Cloud Run means running the backend logic that orchestrates model calls, manages state, and integrates with tools and APIs. An agent receives a user request at its Cloud Run endpoint, calls an AI model such as Gemini on Vertex AI, and then calls tools, APIs, or data sources before returning an answer. This separation keeps the agent logic in the customer-controlled container while the model runs as a managed capability.
Three patterns support agent development on Cloud Run:
Securing agents and supporting other AI workloads on Cloud Run extends the same container and networking controls to agents, retrieval systems, and APIs. Cloud Run integrates with Agent Platform to provide built-in capabilities for AI agents, MCP servers, and tools. That integration helps organizations secure agent deployments without giving up control over how services are packaged, connected, and scaled.
Beyond agents, Cloud Run supports several other AI workloads:
Data Privacy, Security, and Governance Protocols are the rules, controls, and services that keep customer data under customer control in Google Cloud. They define who may access data, how data is encrypted, how violations are detected, and which independent standards verify the controls. A central commitment is that data stored on Google systems belongs to the customer and is not used to train foundation models without permission.
Data ownership and AI training limits define what Google may and may not do with customer data. Data stored on Google systems belongs to the customer, and Google does not scan it for advertising, sell it to third parties, or use it to train AI models without permission. The Data Processing Addendum for Google Cloud states that Google will not process data for any purpose other than meeting contractual obligations. If use of the services stops, tools allow data to be taken elsewhere without penalty or additional cost.
Administrative access and transparency controls determine when Google employees can reach customer data and how that access is recorded. Google logically isolates each customer's data from other customers even when data shares the same physical server. Only a small group of employees can access customer data, with access rights based on job function and role using least privilege and need-to-know. Requests for additional access follow a formal process with approval from the data or system owner, manager, or other executives, while workflow tools keep audit records and security, privacy, and internal audit teams monitor access.
Access Approval (opens in a new tab) and Access Transparency work together to control and record support access. Access Transparency logs administrative access by authorized Google personnel when they support an organization or maintain service availability. Access Approval requires customer approval before an authorized Google administrator can access data for support, and approved requests are logged with Access Transparency logs linked to the approval request. After approval, access must still be properly privileged within Google before it is permitted, and support is provided only to authorized customer administrators.
Exam tip: Access Transparency records administrative access after it happens, while Access Approval requires explicit customer approval before an authorized Google administrator can access data.
Encryption and isolation controls protect customer data at rest, in transit, and in use while keeping keys and data separated when required. Cloud KMS with Cloud EKM (opens in a new tab) lets data be encrypted with keys stored and managed in a third-party key management system outside Google infrastructure, so data at rest stays separate from keys while cloud compute and analytics can still be used. Confidential Computing encrypts data in use during processing. BigQuery automatically encrypts all data at rest and in transit while allowing customized encryption settings.
Network isolation adds defense in depth around the data. Only authorized services and protocols that meet security requirements may traverse Google's network, while firewalls and access control lists enforce network segregation. Traffic is routed through GFE servers to detect and stop malicious requests and distributed denial-of-service (DDoS) attacks. Logs are examined for exploitation of programming errors, and access to networked devices is restricted to authorized employees.
The incident detection and response cycle is the ordered process Google uses to find and handle suspected access violations. It combines expert investigation teams with manual and automated controls that use machine learning, data processing pipelines, and threat intelligence. Each stage feeds the next, so signals become detections, detections become investigations, and investigations improve future protections.

A left-to-right process flow showing Signal Development to Threat Detection to Incident Investigation to Remediation, with a feedback loop returning new threat indicators to the start.
Compliance and governance frameworks are independent certifications, attestations, and audit reports that verify Google Cloud 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, with SOC 2 and SOC 3 reports available to customers. Sector and country programs include FedRAMP for the US government, BSI C5 for Germany, and MTCS for Singapore.
Assured Workloads (opens in a new tab) helps enterprises that need dedicated regional support, certified programs such as FedRAMP or ITAR, or programs such as Sovereign Controls for EU. It provides an enablement workflow to create and monitor the lifespan of required control packages. For law enforcement requests, the data owner is primarily responsible for responding, Google reviews direct government and court requests for legal validity, requests generally must be in writing under appropriate law and signed by an authorized official, and Google notifies the customer unless prohibited by law or court order. Third-party suppliers are assessed for security and privacy before onboarding and must accept security, confidentiality, and privacy contract terms.
Services that control Google access to data give customers technical ways to require approval, record access, enforce regulatory controls, manage keys, or protect data in use. Each service sits between Google administrators or infrastructure and customer data, so support flows from a Google request to customer control to logged access.
| Google Cloud service | Control it provides |
|---|---|
| Access Approval | Requires customer approval before an authorized Google administrator can access data for support, with approved requests logged and linked to Access Transparency logs |
| Access Transparency | Logs administrative access by authorized Google personnel during support or maintenance of service availability |
| Assured Workloads | Provides workflow to create and monitor control packages for regional support and regulatory programs such as FedRAMP, ITAR, or Sovereign Controls for EU |
| Cloud KMS with Cloud EKM | Encrypts data with keys managed outside Google infrastructure while still allowing cloud compute and analytics |
| Confidential Computing | Encrypts data in use during processing |
Data governance and sharing in BigQuery (opens in a new tab) applies privacy, security, and governance principles to analytics without losing control of the data. BigQuery audit logs record user activity and system events to enforce governance policies and identify risks. Data masking obscures sensitive values while allowing access to surrounding data. Access controls include predefined roles, custom roles, row-level and column-level controls, VPC Service Controls perimeters, and Sensitive Data Protection scans for personally identifiable information (PII), financial data, and health information.
BigQuery sharing allows data owners to share insights within and across organizational boundaries while retaining control. Shareable assets include datasets, tables, views, real-time streams with Pub/Sub topics, SQL stored procedures, and BigQuery ML models. Data owners define rules or policies that restrict access and usage. Live, zero-copy sharing shares data in place with no movement or replication through linked datasets that act as live pointers to the shared asset. Publishers can monitor usage metrics such as jobs executed, total bytes scanned, and subscribers, and data clean rooms allow collaboration on sensitive data.
Model Ecosystem and Selection Flexibility is the ability to choose among leading first-party models and open-source alternatives based on performance, cost, and transparency needs while keeping data under customer control. A diverse ecosystem avoids locking users into a single model and lets different workloads use different models. Google Cloud applies the same security, privacy, and governance protections regardless of which model is selected.
Why model choice gives control is that different models offer different tradeoffs, so the workload determines the selection instead of platform limits. A team can select a high-capability first-party model when accuracy and managed features matter most, or select an open model when inspection, portability, and customization matter most. This flexibility ties the decision to performance, cost, and transparency needs while the same data protections continue to apply underneath.
How customer data stays under customer control is through ownership rules, contractual limits, and infrastructure isolation that apply behind any model. Stored customer data belongs to the customer, and Google Cloud does not scan it for advertising, sell it to third parties, or use it to train AI models without permission. The Data Processing Addendum states that Google will not process data for any purpose other than meeting contractual obligations, and tools support moving data elsewhere without penalty or additional cost when service use stops.
Administrative isolation reinforces ownership on shared infrastructure. Each customer's data is logically isolated from other customers even on the same physical server, and only a small group of employees can access customer data. Access follows least privilege and need-to-know tied to job function and role, with extra access requiring a formal request and approval workflow that maintains audit records. Dedicated security, privacy, and internal audit teams monitor employee access, with visibility provided through logging and approval controls.
Services that monitor and restrict support access give users visibility into Google administrator activity and technical means to limit it. These services sit between Google administrators and customer data, so a support request flows from Google personnel to customer approval to logged, privileged access.
Exam tip: Access Transparency logs Google administrator access, while Access Approval requires customer approval before that access is permitted.
The incident response cycle behind any model choice protects data through a continuous loop from signals to detection to investigation to improvement. Threat intelligence from past incidents, network traffic, access logs, anomalous behavior, and offensive security exercises feeds detection systems that monitor internal systems, alert staff, and trigger automated responses such as revoking access to a resource. The same cycle applies no matter which first-party or open model the workload uses.
Access Transparency logs administrative access by authorized Google personnel when they support an organization or maintain service availability, while 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 request, and access must still be properly privileged within Google before it is permitted.
Cloud Run hosts AI agents, chatbots, and other AI workloads in containers that expose a secure Cloud Run URL endpoint over HTTP with streaming support and connect to services such as Vertex AI, BigQuery, Cloud SQL, Memorystore, Pub/Sub, AlloyDB for PostgreSQL, Cloud CDN, and Secret Manager. It offers flexible container packaging, automatic scaling based on demand or manual scaling, GPU support that can scale down to zero when idle, direct VPC connectivity, and granular security and networking controls. Beyond agents, it supports serving AI/ML models for inference, building Retrieval-Augmented Generation (RAG) systems, hosting AI-powered APIs and backends, and prototyping new AI ideas.
Data stored on Google systems belongs to the customer, and Google does not scan it for advertising, sell it to third parties, or use it to train AI models without permission. The Data Processing Addendum for Google Cloud states that Google will not process data for any purpose other than meeting contractual obligations, and tools allow data to be taken elsewhere without penalty or additional cost if use of the services stops.
An enterprise is designing a corporate generative AI strategy across multiple departments with differing requirements:
What is the primary strategic advantage of leveraging a diverse model ecosystem that combines first-party foundation models and open-source alternatives?
Last updated