Microsoft AI-300 Exam Prep
Operationalizing Machine Learning and Generative AI Solutions (Page 6 )

• Input feature distributions differ from training data ? Analyze dataset drift metrics

• Model accuracy drops without code changes ? Analyze dataset drift metrics

• Endpoint latency increases under load ? Investigate scaling and infrastructure metrics

• Accept: Acknowledge the risk and choose not to take additional action.
• Transfer: Shift the risk or its financial impact to another party, such as through insurance or outsourcing.
• Mitigate: Reduce the likelihood or impact of the risk, such as by implementing security controls.
• Avoid: Eliminate the activity or condition that creates the risk.

• Multiple nodes: Data is replicated several times in the primary region.
• Separate geography: A copy is asynchronously replicated to a secondary Azure region.
• Readable secondary: RA-GRS provides read-only access to the secondary replica, even when the primary region is available.

• GRS: Replicates to another region but the secondary copy isn’t normally readable.
• ZRS: Replicates across availability zones in the same region, not a separate geographic region.
• LRS: Keeps replicas within a single datacenter or region and provides no geographic redundancy.

• Encryption to preserve confidentiality
• Integrity protection to detect alteration
• Authentication of the server through certificates

• SHA-256: A hashing algorithm; it does not encrypt data for transmission.
• SSL 3.0: Obsolete and insecure. It should not be used.
• AES-256: A strong encryption algorithm, but by itself it does not provide the complete secure communication protocol needed for data in transit.

1. Get Data ? Folder in Power BI Desktop.
2. In Power Query Editor, expand the Attributes column to expose metadata such as hidden status and other file attributes.
3. Remove the Content column (and any other data columns you do not need).

• Name
• Extension
• Date accessed
• Date modified
• Folder Path
• Attributes
• Content

• Manage multiple organizations from one hierarchy.
• Enterprise-wide policies, including repository visibility, forking, and GitHub Copilot policies.
• Centralized identity and access management, such as SAML SSO, SCIM provisioning, and—where applicable—Enterprise Managed Users.
• Centralized billing and licensing, including organization and Copilot seat management.
• Enterprise audit log, which aggregates activity across organizations.
• Security and compliance oversight, such as organization security visibility and enterprise-wide controls.
• Enterprise support and account management features, depending on the plan.

• Organization: manages its own repositories, teams, members, and local policies.
• Enterprise: governs multiple organizations and can enforce policies across them.

• SAML: Enables federated authentication and single sign-on between the IdP and SaaS providers.
• OAuth tokens: Allow delegated application access; OAuth is not primarily the first step for employee SSO.
• Password vaulting: Stores or autofills multiple passwords but does not eliminate the separate credentials or provide true federation.

• SQL injection
• Cross-site scripting (XSS)
• Malicious file uploads
• Common web-application exploits

• Layer 4 firewall: Filters primarily by IP addresses, ports, and transport protocols; it lacks detailed web-application inspection.
• NGFW: Provides advanced network filtering and may include application awareness, but a WAF is more specialized for protecting web applications.
• UTM: Combines several security functions in one device, but it is not specifically focused on web-application attacks.

• Review iteration length: Changing duration does not resolve unclear requirements.
• Limit work in progress: Useful when excessive multitasking is the problem, but it does not directly clarify requirements.
• Review individual performance: The issue is a process and backlog-quality problem, not necessarily an individual performance problem.

1. Ensure BOMs can include subcomponents ? Configure an entity relationship

1. Report changes made by engineers ? Enable auditing

• Configure entity relationship
• Configure auditing

• The user authenticates with the external Identity Provider.
• Salesforce receives the SAML assertion.
• Salesforce uses attributes in that assertion—such as username, email, first name, last name, and profile/community information—to create the user during the first login.
• On later logins, Salesforce updates or uses the existing user record.

• On-the-fly provisioning is a general phrase; the Salesforce capability being tested is specifically JIT provisioning.
• Salesforce APIs would require a separate integration and provisioning process.
• Identity Connect is designed primarily to synchronize users from an identity directory, not to create users at the moment of their first SAML login.

• Discovers sensitive data in Amazon S3 buckets
• Uses machine learning and pattern matching to identify data such as personally identifiable information (PII)
• Helps assess S3 security and privacy risks
• Generates findings for potentially sensitive or exposed data

• Amazon Detective: Investigates security findings and performs root-cause analysis.
• Amazon GuardDuty: Detects threats and suspicious activity across AWS accounts and workloads.
• AWS IAM Access Analyzer: Identifies resources that are accessible from outside their intended trust zone, such as publicly accessible S3 buckets.

• vSmart controls the overlay control plane. It uses OMP to exchange routing information, distribute policies, and determine how WAN Edge devices reach one another.
• vManage is the centralized management and configuration platform.
• vBond is the orchestrator. It helps devices authenticate, discover controllers, and traverse NAT during onboarding.
• APIC-EM is not the SD-WAN control-plane component described here.

• Interactive development and exploration
• Small datasets
• Lightweight transformations or testing
• Avoiding the overhead and cost of multiple workers

• Automated reports: Jobs may benefit from job clusters or SQL warehouses.
• Working with SQL in Databricks SQL: Databricks SQL warehouses are the relevant compute option.
• Automatically scaling for larger data: A multi-node cluster with autoscaling is more suitable.
• Large amounts of data: Distributed worker nodes are generally needed.

• A single Construction Job can use multiple Machinery records.
• A single Machinery record may be used on multiple Construction Jobs.
• This is a many-to-many relationship, which Salesforce models using a junction object.
• The junction object can contain fields such as:

• Master-detail to Machinery
• Master-detail to Construction Job
• Date and cost fields

• Mesh: WAPs communicate wirelessly with one another, often to extend coverage without wired connections. It does not specifically describe centralized WLC management.
• Point-to-point: Connects two specific devices or locations.
• Ad hoc: Wireless devices communicate directly with one another without an access point.

• R-squared (R²): mede quanto da variação dos valores reais é explicada pelo modelo. É mais usado em regressão, não em classificação.

• Accuracy (acurácia): mede a proporção de previsões corretas em relação ao total. É adequada para a classificação de imagens da questão.

• RMSE (Root Mean Squared Error): mede o tamanho médio dos erros entre valores previstos e reais, penalizando erros maiores. É usado principalmente em regressão.

• Learning rate (taxa de aprendizado): controla o tamanho dos ajustes feitos pelo modelo durante o treinamento. Não é uma métrica de avaliação; é um hiperparâmetro.

• Valor maior: mais aderência às instruções e aos detalhes descritos.
• Valor menor: mais liberdade criativa e resultados potencialmente mais variados.
• Valor excessivamente alto: pode causar artefatos ou imagens menos naturais.

• Aumentar o número de passos de geração: pode melhorar qualidade e refinamento, mas não é o principal controle de especificidade do prompt.
• MASK_IMAGE_BLACK: relaciona-se ao uso de máscara em geração ou edição de imagens, não à especificidade textual.
• Prompt strength: é mais associado à força do prompt em processos como image-to-image; nesta questão, o parâmetro diretamente ligado à aderência textual é o CFG scale.

1. Convert text and images into numerical embeddings.
2. Store or compare those embeddings.
3. Retrieve the most semantically similar items.

• Embedding model: Best for similarity search and retrieval.
• Generation model: Best for producing new text or other content from multimodal input.

• Date: 2018-12-31
• Time: 08:59

• B: The original value contains the text at, so it cannot be converted directly to a date.
• A: The last characters contain the time, not the date.
• D: Extracting the first 11 characters can leave an unwanted space and does not properly separate the time component.

Question 3: AWS Storage Gateway File Gateway

• A: S3 bucket per user — S3 is object storage, not a traditional shared file system; mounting it requires extra tools and is less efficient.
• B: File Gateway — Provides a local file share using standard file protocols while storing files in Amazon S3. This preserves convenient local access and extends storage capacity.
• C: WorkSpaces and WorkDocs — Designed for virtual desktops and document collaboration, not large shared file storage.
• D: EC2 with EBS — EBS is attached to one EC2 instance and is not designed for direct multi-user file sharing.

• Cloud architecture
• Data architecture
• Platform engineering
• Data engineering
• Provisioning and modernization

• Data protection is associated mainly with security and risk management.
• Data governance belongs to the governance perspective.
• Data science is not an AWS CAF capability.

• Business
• People
• Governance
• Platform
• Security
• Operations

• Program and project management
• Risk management
• Cloud financial management
• Data governance
• Portfolio management

• Sustainability
• Performance efficiency
• Reliability

• Primary and core nodes: Use On-Demand Instances because the workload is business-critical. These nodes manage the cluster and store/process HDFS data.
• Task nodes: Use Spot Instances in an instance fleet. Task nodes are suitable for interruption because they do not hold essential cluster data and can be replaced or diversified across instance types.
• Terminate the entire cluster after processing: The job runs only about six hours each morning, and EMRFS stores data in Amazon S3. Keeping the cluster running for the rest of the day would waste money.

• A: Using Spot for primary and core nodes risks interruption of the critical cluster.
• C: All On-Demand is reliable but unnecessarily expensive; terminating the cluster helps, but Spot task nodes provide greater savings.
• D: Leaving primary and core nodes running after the task finishes continues to incur costs. Savings Plans also are less appropriate for this intermittent workload.

• A. Tuning system software: This is a corrective action, not an analysis method. Tuning should be based on measured performance data.
• B. User dissatisfaction reports: Useful as an indicator, but subjective and insufficient for diagnosing the cause.
• D. Off-peak utilization and response time: This may hide problems occurring during peak periods. Performance analysis should cover representative or peak workloads.

• Secret scope: A named container that groups related secrets and defines access permissions. Example: finance-prod.
• Secret key: The name of an individual secret stored inside that scope. Example: db-password.

text 
Scope: finance-prod 
  Key: db-username 
  Key: db-password

python 
dbutils.secrets.get( 
    scope="finance-prod", 
    key="db-password" 
)

• B. Users accessing Salesforce from public Wi-Fi

• C. Users choosing the same password as their Facebook password

• A: 2FA does not protect an already-authenticated, unlocked session if someone accesses an unattended laptop.
• D: Weak password-reset questions are an account-recovery weakness. 2FA primarily strengthens login authentication, not necessarily the security-question reset process.

• Reviewing and validating generated code
• Checking security, licensing, and quality concerns
• Testing the implementation
• Making the final design and coding decisions

• A: Copilot is not a complete replacement for developers.
• C: Human input, review, and oversight are still required.
• D: Copilot Chat supports many tasks beyond debugging, including explanation, generation, refactoring, and testing assistance.

• Updating the CRL: Revokes compromised certificates, but does not fix the CA flaw that allowed them to be created.
• Changing passwords: May be necessary if accounts were compromised, but it does not address the certificate-generation vulnerability.
• Implementing SOAR: Improves security automation and response; it does not remediate the underlying PKI weakness.

• Snapshot: Captures a point-in-time state for rollback. It is not an independent running VM and does not protect against failure of the underlying host or storage.
• Convert to dynamic disks: Changes disk allocation behavior; it does not provide redundancy or high availability.
• P2V: Converts a physical machine into a virtual machine. That is a migration process, not an availability solution.

• An Azure pay-as-you-go subscription follows an operational expenditure (OpEx) model.
• With OpEx, the organization pays for services as they are used, typically based on consumption.
• There is no need to purchase and maintain physical servers upfront.
• This contrasts with capital expenditure (CapEx), where the company makes a significant upfront investment in infrastructure, such as on-premises Hyper-V hosts.

• Disable default accounts: Default or unused accounts may have predictable names or credentials and can be abused by attackers. They should be disabled, removed, or secured with strong credentials where necessary.
• Remove unnecessary services: Every running service can contain vulnerabilities or provide an entry point. Removing services the server does not need reduces exposure.

• Add the server to the asset inventory: Important for asset management, but it does not directly harden the system.
• Document default passwords: Default passwords should be changed or disabled—not recorded for continued use.
• Send logs to the SIEM: Supports monitoring and detection, rather than directly reducing the attack surface.
• Join the corporate domain: A deployment/configuration decision, not inherently a hardening measure.

Why

cmd 
PowerShell.exe -ExecutionPolicy Bypass -File C:\Scripts\ConfigureVM.ps1

Why the other options are less suitable

• GPO logon script: Runs only when a user logs on and may not run during initial VM provisioning.
• GPO startup script: Can work, but requires the VM to be joined to the domain and the policy to be applied. It is also normally evaluated at each startup.
• Place scripts in a VHD: Merely placing scripts on a disk does not cause them to execute.

• Data loss prevention (DLP): Helps detect or prevent unauthorized data transfers, but does not directly protect data on a stolen, powered-off laptop.
• Operating system hardening: Reduces vulnerabilities and attack surface, but does not adequately protect stored data if the device is stolen.
• Boot security: Helps prevent unauthorized operating systems or boot media from loading. It supports device security but is not the primary solution for protecting the data itself.

• Windows 10 VMs
• Windows Server 2012 or later VMs
• Running VMs—they do not need to be shut down
• Debian 8.2+ VMs, assuming the specific version is supported
• Stopped or deallocated VMs

• Conserve public IPv4 addresses: IPv4 addresses are limited, so many internal devices share a smaller number of public addresses through NAT.
• Reduce direct exposure: Private addresses are not directly routable from the public Internet, which helps reduce unsolicited access to internal devices.

• prompt: the input or instruction given to the model
• completion: the desired response

json 
{"prompt":"Summarize this customer complaint.","completion":"The customer reports a delayed shipment."}

• B: A .txt file containing CSV-style lines is not automatically a valid fine-tuning dataset. The required structure and format must match the selected Bedrock model.
• C: Provisioned Throughput provides dedicated inference capacity; it does not fine-tune a model.
• D: Journals and textbooks might support domain adaptation or pretraining, but they do not provide the prompt-to-answer examples typically required for supervised fine-tuning.

• The inference container image or built-in algorithm
• The location of model artifacts in Amazon S3
• IAM permissions through the execution role
• Optional environment variables and networking settings

1. Create AWS::SageMaker::Model
2. Create an AWS::SageMaker::EndpointConfig that references the model
3. Create AWS::SageMaker::Endpoint that references the endpoint configuration

• AWS::SageMaker::Endpoint hosts a model but does not define the model itself.
• AWS::SageMaker::NotebookInstance provides a development environment.
• AWS::SageMaker::Pipeline orchestrates ML workflow steps.

• A. Configure prompt modifications to enforce tone, disclaimers, and refusal behavior at the system level.

• D. Configure Power Platform DLP policies to restrict unauthorized data connectors.

• B: Manual topic-by-topic disclaimer edits are specifically discouraged.
• C: Channel settings do not replace agent-level Responsible AI controls.
• E: Publishing before enabling safety filters violates the stated deployment sequence. Responsible AI filters should be enabled before publication, not afterward.

• Drops packets when traffic exceeds a certain bitrate: This describes traffic policing, not shaping.
• Fragments large packets and queues them: Shaping does not fragment packets as its primary congestion-control action.
• Drops packets randomly from lower-priority queues: Packet dropping may occur if a shaping queue becomes full, but it is not the defining behavior.

• Shaping: buffers and delays excess traffic.
• Policing: drops or marks traffic that exceeds the configured rate.

• Use private connectivity from a cloud provider
• Set up SSO for federated authentication

• Private connectivity—such as a cloud provider’s private endpoint—keeps Snowflake access within the company’s private network path instead of exposing access over the public internet. This fits the VPN/VDI requirement.
• SSO for federated authentication allows Snowflake users to authenticate using the credentials already managed by the company’s identity provider—the same identity system used for VDI—eliminating separate Snowflake passwords.

• Tri-Secret Secure protects encryption keys; it does not provide private network access or credential reuse.
• A proxy Snowflake account outside the VPN conflicts with the requirement that Snowflake have no internet visibility and is not the normal solution for federated login reuse.

• A — Can help keep costs down and improve quality: The provider may have specialist expertise, efficient systems and economies of scale.
• B — There is a known fee if agreed in advance: A fixed or agreed fee can make costs more predictable and support budgeting.
• C — Allows management to focus on core operations and issues: Senior managers spend less time supervising routine finance processes and more time on strategic matters.

• Increased control: Control may decrease because the activity is performed externally.
• Increased independence: The organisation becomes dependent on the supplier.
• Increased understanding of the outsourcing business: The internal organisation usually has greater business knowledge; the provider may initially lack it.
• Increased confidentiality: Sharing financial information externally may create confidentiality risks.
• Internal expertise improved: Outsourcing can reduce opportunities to develop internal finance skills.

1. Start with the Customer table.
2. Use Merge Queries and select the Address table.
3. Match Customer[Address ID] to Address[Address ID].
4. Use a Left Outer join so every customer is retained.
5. Expand only City, State/Region, and Country.

• Append adds rows vertically; it is for stacking similarly structured tables.
• Group By summarizes rows and does not retrieve address columns.
• Transpose swaps rows and columns and is unrelated to table relationships.

• Works across SaaS and internal applications: TOTP is widely supported by authentication systems and can be integrated into many applications.
• Device manufacturer agnostic: A hardware token or authenticator application can generate the code without depending on a specific computer or phone manufacturer.
• Offline capability: The code is generated locally using a shared secret and the current time; an internet connection or cellular service is not required.

• Username and password: May work offline but provides weaker authentication and does not meet the intended improvement in security.
• Biometrics: Can depend on device hardware, operating-system support, and enrollment mechanisms.
• SMS verification: Requires cellular connectivity and is vulnerable to threats such as SIM swapping.

• A. Microsoft 365 group with Assigned membership
• C. Microsoft 365 group with Dynamic User membership

Why?

• Assigned membership: You explicitly add User1, User2, and User3.
• Dynamic User membership: Membership is automatically determined by a rule based on user attributes.

Why the other options are incorrect

• B and D: Security groups do not support the Microsoft 365 group expiration policy.
• E: Dynamic Device groups are intended for devices, not these users, and are also security groups.

• ARP poisoning: Redirects or intercepts local network traffic; monitoring would typically show suspicious ARP changes or man-in-the-middle behavior.
• Jamming: Interferes with wireless or radio signals. It would affect connectivity, not usually show a server-specific overload.
• Kerberoasting: Attempts to obtain and crack Kerberos service-account hashes; it does not normally cause a system outage.

• Vulnerability scanner: Finds technical weaknesses in systems, applications, or devices. It does not usually analyze the entire supply chain.
• Penetration test: Simulates attacks against a defined target to identify exploitable weaknesses. Its scope is narrower.
• SCAP (Security Content Automation Protocol): Provides standardized formats and methods for automating security configuration and vulnerability checks, rather than mapping the full supply chain.

• Progress iteratively with feedback focuses on making improvements in manageable steps and learning from feedback.
• Optimize and automate focuses on maximizing value from resources and automating suitable activities.
• Start where you are focuses on assessing the current situation and reusing what already exists.

• Secure: uses encryption, such as HTTPS or SFTP.
• Cross-platform: supported by many languages and operating systems.
• Internet-friendly: designed to communicate across networks, usually using standard Internet protocols.

• JSON/REST over HTTPS — widely supported, secure, and Internet-friendly.
• SOAP over HTTPS — also secure and highly cross-platform, though more heavyweight.
• XML over SFTP — secure file-based exchange, but it is not an API-style interface.

• Dynamics 365 Remote Assist — used for remote collaboration and mixed-reality scenarios.
• Dynamics 365 Finance — supports financial and operational management.
• Dynamics 365 Business Central — designed primarily for small and medium-sized businesses’ finance and operations.
• Dynamics 365 for Phones — provides mobile access to Dynamics 365 apps, including Sales, through the mobile Unified Interface.