Prepare and test your skills
Prepare and test your skills
Infrastructure hardening and inference protection is the securing of the environment around machine learning models and the APIs that serve predictions. Gemini Enterprise Agent Platform (opens in a new tab) provides a secure and governed environment to develop and deploy models, and it applies Google Cloud security controls to protect models and training data. Hardening adds network perimeters, edge filtering, and supply-chain checks so deployments stay isolated and inference endpoints resist web exploits, distributed denial-of-service (DDoS) attacks, prompt injection, and unauthorized model extraction.
Isolating model deployments with network boundaries means placing models, retrieval systems, and indexes inside a controlled perimeter that blocks outside access by default. VPC Service Controls (VPC-SC) (opens in a new tab) creates that perimeter around Google Cloud resources, so a model deployment, retrieval engine, or vector index inside the boundary rejects requests from outside unless they are explicitly allowed. Secure API endpoints extend this isolation by keeping inference traffic inside trusted boundaries and exposing it only through controlled and monitored paths.
The set of perimeter and data protections varies by machine learning feature:
| Feature | Data residency at-rest | Customer-managed encryption key (CMEK) | VPC Service Controls (VPC-SC) | Access Transparency (AXT) |
|---|---|---|---|---|
| Gemini Enterprise Agent Platform | Yes | Yes | Yes | Yes |
| RAG Engine | No | No | Yes | No |
| Vector Search | Yes | No | Yes | No |
RAG Engine supports only VPC-SC, while Vector Search supports data residency and VPC-SC. Vertex AI Feature Store and Vertex Data Labeling do not meet data-at-rest commitments, so location guarantees do not apply to them in the same way.
Exam tip: Choose Gemini Enterprise Agent Platform when you need all four controls, but remember RAG Engine supports only VPC-SC.
Protecting inference APIs at the network edge means filtering hostile internet traffic before it reaches model serving systems. Google Cloud Armor (opens in a new tab) provides that edge defense against web exploits and DDoS attacks. Traffic flows from clients on the internet to Google Cloud Armor first at the edge, and only filtered traffic then flows onward to the web application or inference API behind it. This placement creates a trust boundary with attackers and automated exploit traffic on the outside and model serving systems on the inside.
A left-to-right process flow showing internet traffic hitting Google Cloud Armor at the edge, with only filtered traffic reaching the inference API inside while attacks are blocked and Security Command Center monitors findings.
Security Command Center aggregates and prioritizes security findings and recommendations so teams can address risks to those endpoints. Its Web Security Scanner automatically scans web applications for common vulnerabilities when integrated into development pipelines. Together, edge filtering plus centralized findings help endpoints stay resilient during attack spikes and reduce exposure that could enable prompt injection or model extraction.
Hardening builds, provenance, and workloads means verifying what code is allowed to run near models and data. The Supply-chain Levels for Software Artifacts (SLSA) (opens in a new tab) framework defines security checks for source code, the build process, and code provenance. Binary Authorization enforces that provenance by allowing only verified workloads to run, and Assured Open Source Software ensures workloads use only verified open-source libraries.
Container and deployment checks extend the same control to running workloads. The container scanning feature of Artifact Registry tracks vulnerabilities that are identified and remediated, including the percentage of deployments that pass scans. Alignment with the Secure AI Framework (SAIF) provides a practical approach for addressing security and risk concerns across the AI lifecycle, from governed development to hardened deployment and monitoring.
Secure data governance and model development is the protection of training data, models, and workflows from ingestion through training and deployment. Gemini Enterprise Agent Platform (opens in a new tab) provides a secure and governed environment for developing and deploying machine learning models. Sensitive Data Protection (opens in a new tab) discovers and masks sensitive data before storage, while network perimeters, identity controls, and authentication lock down who can reach data and training resources.
A secure and governed training environment is a machine learning platform where models are developed, evaluated, and deployed under consistent security and responsibility controls. Gemini Enterprise Agent Platform implements Google Cloud security controls to help secure models and training data, so the training environment inherits protections for data integrity and operational resilience. Built-in capabilities such as model explainability and fairness metrics help address responsible-AI concerns during development and deployment. Teams strengthen this foundation with clear governance policies for AI development and deployment and by aligning with the Secure AI Framework (SAIF), which provides a practical approach to AI security and risk.
Protecting sensitive data during ingestion is the process of finding confidential information in incoming data and masking it before it reaches storage or training. Confidential data includes credit card numbers, addresses, phone numbers, and other personally identifiable information (PII). Sensitive Data Protection identifies, classifies, and protects this data, and Dataplex Catalog stores, manages, and provides access to metadata about it. Data flows from the ingestion source through Sensitive Data Protection for classification and masking, and only then to Google Cloud storage and the model training environment.
Sensitive data protection follows an ordered flow:
Only masked values move forward to storage and training when this order is followed.
Locking down access means placing sensitive data and machine learning resources behind network boundaries and allowing only verified identities to reach them. VPC Service Controls (VPC-SC) (opens in a new tab) creates a security service perimeter around Google-managed services, with sensitive data inside the perimeter and untrusted networks and unauthorized projects outside it. The perimeter controls movement of data across its boundary, which improves the ability to mitigate data exfiltration, the unauthorized copying or transferring of data. For regulated workloads, Assured Workloads sets up data boundaries that help meet government and industry requirements such as FedRAMP and CJIS.
Identity controls decide who can cross that boundary and under what conditions. Identity and Access Management (IAM) applies access controls to data inside the perimeter, so only granted roles and service accounts can read training datasets or manage models. Multifactor authentication (MFA) is configured for all users who require access to sensitive data, so a password alone is not enough to enter the training environment. Together, the VPC-SC perimeter depends on IAM and MFA to enforce both where data can move and who can move it.
Security controls available to machine learning services differ by feature, so not every feature inherits the same protections. Gemini Enterprise Agent Platform supports the broadest set of controls, while RAG Engine and Vector Search support a smaller subset focused on perimeter isolation and residency.
| Feature | Data residency at-rest | Customer-managed encryption key (CMEK) | VPC Service Controls | Access Transparency |
|---|---|---|---|---|
| Gemini Enterprise Agent Platform | Yes | Yes | Yes | Yes |
| RAG Engine | No | No | Yes | No |
| Vector Search | Yes | No | Yes | No |
This comparison matters because a workflow that depends on RAG Engine cannot rely on CMEK or Access Transparency and must rely on VPC-SC for isolation. A workflow that depends on Vector Search keeps data residency and perimeter protection but still lacks customer-managed key control.
Exam tip: Vertex AI Feature Store and Vertex Data Labeling do not meet data-at-rest commitments, even when other platform features do.
Continuous monitoring and threat detection is the practice of watching an AI system's logs, findings, and performance signals throughout its operational life. Logging services record what happened, detection services analyze those records for attacks and misconfigurations, and response services act on what they find. This loop catches attacks, misconfigurations, and model drift early instead of after damage occurs.
Audit logging and visibility is the foundation of monitoring because it records who did what, where, and when inside Google Cloud and AI resources in real time. Google Cloud separates this visibility into three log types that answer different questions.
| Log type | What it records | Collection |
|---|---|---|
| Platform logs | Activity within the AI service and its Google Cloud resources | Always collected |
| Audit logs | The who did what, where, and when details of resource access | You must enable them |
| Access Transparency logs (AXT) | Actions taken by Google personnel on your resources | Records Google-side access |
Model monitoring is separate from logging. Logs show who touched a system, while model monitoring shows whether the model itself still behaves correctly. Some services, such as Anti Money Laundering AI (AML AI) (opens in a new tab), provide artifacts that support model risk governance processes, which a team can feed into its own monitoring workflow.
Exam tip: Platform logs are always collected, but audit logs are not — you must turn audit logging on to collect them.
Detecting vulnerabilities across the machine learning lifecycle means scanning the artifacts an AI system is built from before and after deployment. Security Command Center (opens in a new tab) is the central place where findings and recommendations are aggregated and prioritized, so teams act on the highest-risk items first. Its Web Security Scanner automatically scans web applications for common vulnerabilities and can be integrated into development pipelines, so scans run as part of the build rather than after release.
Supply-chain checks protect the code and models that enter the environment:
These controls connect in sequence: a scanner finds a vulnerability, the finding flows into Security Command Center for prioritization, and Binary Authorization blocks any workload whose provenance does not pass verification.
Detecting anomalies and responding to threats uses analytics and automation to find attacks that fixed rules alone would miss. Google Security Operations analyzes large volumes of security data with AI and machine learning, detects anomalies, and automates threat response. Its SOAR and SIEM dashboards supply the metrics used to judge how well detection and response are working. Mandiant services add proactive defense through threat intelligence, incident response, and security validation, which help identify and mitigate potential threats before they cause damage.
Protection at the network edge is handled by Google Cloud Armor, which defends web applications and APIs against web exploits and DDoS attacks. Teams should also validate defenses with regular security simulations and penetration testing, because a detection rule that has never been tested may not fire during a real incident. Security Command Center sits across these tools as the aggregator that turns scattered findings into a prioritized list of risks.
Monitoring model performance and drift means watching the model as well as the infrastructure around it. Gemini Enterprise Agent Platform provides a secure and governed environment for developing and deploying models, including model explainability and fairness metrics that support responsible-AI review. Aligning with the Secure AI Framework (SAIF) (opens in a new tab) gives teams a practical structure for addressing security and risk concerns, and clear governance policies for AI development and deployment should cover both security and ethics.
Data handling supports this monitoring. Assured Workloads sets data boundaries for sensitive and regulated workloads, and Sensitive Data Protection identifies, classifies, and protects sensitive data such as financial information. Identity and Access Management (IAM) grants granular access to specific AI and Google Cloud resources, while VPC Service Controls adds a perimeter on top of IAM that blocks access from unauthorized networks and services outside the perimeter. Data residency policies ensure data is stored only in the regions specified when an AI instance is created, and services such as AML AI support regional but not multi-regional deployment.
Metrics that show monitoring works are the indicators that prove detection and response are actually improving. These indicators come from Google Security Operations SOAR and SIEM dashboards and from tracking tools used during development:
Together these metrics close the monitoring loop: logs and scans produce findings, detection services turn findings into alerts, response services act on the alerts, and the metrics show whether the next cycle detects threats faster and more accurately than the last.
VPC Service Controls (VPC-SC) creates a security perimeter around Google Cloud resources so a model deployment, retrieval engine, or vector index inside the boundary rejects requests from outside unless explicitly allowed. The perimeter controls movement of data across its boundary, which improves the ability to mitigate data exfiltration. Identity and Access Management and multifactor authentication enforce who can cross that boundary and under what conditions.
Gemini Enterprise Agent Platform supports data residency at-rest, customer-managed encryption key (CMEK), VPC Service Controls (VPC-SC), and Access Transparency (AXT). RAG Engine supports only VPC Service Controls (VPC-SC), while Vector Search supports data residency at-rest and VPC Service Controls (VPC-SC). Vertex AI Feature Store and Vertex Data Labeling do not meet data-at-rest commitments, so location guarantees do not apply to them in the same way.
Google Cloud Armor provides edge defense against web exploits and distributed denial-of-service (DDoS) attacks. Traffic flows from clients on the internet to Google Cloud Armor first at the edge, and only filtered traffic then flows onward to the web application or inference API behind it.
Platform logs record activity within the AI service and its Google Cloud resources and are always collected. Audit logs record the who did what, where, and when details of resource access and you must enable them to collect them. Access Transparency logs (AXT) record actions taken by Google personnel on your resources.
A financial services organization is preparing customer service interactions and transaction notes to train and fine-tune a generative AI model on Google Cloud. The project requires meeting strict data governance standards:
Which strategy best meets these security and data governance requirements?
Last updated