Microsoft GH-600 Exam Actual Questions
Developing in Agentic AI Systems (Page 2 )

• Box 1: Azure Lighthouse — ??? ??????? ? ?????????? ????????? ? ?????????? Entra tenants. ??? ??????????? ??? ??? ????? ????? ? VM ? ?????? ?????????/????????.
• Box 2: Azure Monitor Agent ? ?????????????? Data Collection Rules (DCR) — AMA ???????? ??????? Windows (????????, Security Logs) ? DCR ??????????, ????? ??????? ???????? ? ???? ?????????? (? ?????? Log Analytics workspace).

• Why: Table.Profile returns a per-column profile that includes statistics such as count, null count, distinct count, and standard deviation. It also provides min/max and average where applicable, which matches the requested metrics (Count, Average, Null count, Distinct count, Standard deviation).
• Syntax: Table.Profile(table as table, optional additionalAggregates as nullable list) as table
• How it helps: After running Table.Profile on your table, you can extract the per-column stats you need and display them in your semantic model or report.

• Two VMs
• Distributed across two separate availability zones

• Cross-account access is typically implemented with an IAM role. You create a role in the target account (where the resources live) and configure a trust policy that allows principals from the source account to assume that role.
• Users in the source account then assume the role (via Switch Role in the console or STS:AssumeRole) and receive temporary credentials with the role’s permissions to access the resources.

• IAM group: Groups exist inside a single account; they don’t grant cross-account access.
• IAM tag: Tags are metadata and do not grant access.
• IAM Access Analyzer: It analyzes and identifies public or cross-account access but does not grant access itself.

• The AWS Support Concierge is a specialized team within Enterprise Support whose primary role is handling billing and account questions and helping with account setup and getting started.
• They are not the source for architecture guidance or development assistance; those services are provided by Technical Account Managers (TAMs) or Solutions Architects as part of Enterprise Support.
• For technical support cases, customers typically interact with AWS Support engineers or the TAMs, not the Concierge. The Concierge may route or coordinate related requests, but they don’t provide the technical resolution.

• What Concierge does: billing and account inquiries, onboarding help, coordination with AWS support.
• What Concierge does not do: provide architecture/design guidance or in-depth technical support case resolutions.

• Ensure that only users who manage the production environment sign in to the Azure portal from a hybrid Entra-joined device and authenticate with MFA.

• Use Microsoft Entra ID Protection (Identity Protection) to enforce MFA enrollment:

• Update the existing Capolicy1 (a Conditional Access policy):

• Identity Protection MFA registration policy ensures users are enrolled in MFA (Box 1 choice).
• CA policy Capolicy1 with aGrant control of “Require MFA” enforces MFA during sign-in (Box 2 choice).
• Both work together with the hybrid-joined device requirement to satisfy the authentication/authorization rules.

• Apply the MFA registration policy to the production-management users.
• Apply the MFA grant to Capolicy1 for those same users/groups.

• TCP is connection-oriented. It establishes a connection before data transfer using a three-way handshake (SYN, SYN-ACK, ACK), enabling reliable delivery, ordering, and flow control.

• UDP is connectionless. It sends datagrams without a prior connection and provides no guarantee of delivery, order, or protection against duplicates; any reliability is up to the application.

• B incorrectly states UDP uses acknowledgment packets; UDP has no built-in ACKs.
• C wrongly claims UDP provides reliable transfer and that TCP is connectionless.
• D misstates framing bits; neither protocol uses the listed TCP/UDP handshake bits in frames.

• Correct answer: Day 3 (Option B)

• Why: In Engagement Studio, when the program is paused, the prospects’ wait times are also paused and do not count down. They resume from the same point when the program restarts. Here, the prospect was at day 3 of a 5-day wait when paused, so they resume on day 3 after the 1-day pause.

• Used for communication on the local link or LAN segment
• Automatically generated on IPv6-enabled interfaces
• Not routed by routers to other networks
• Commonly used by devices to communicate with their local IPv6 router
• Begins with FE80 through FEBF

• FC00::/7 — unique local unicast
• FE80::/10 — link-local unicast
• FF00::/8 — multicast
• 2000::/3 — global unicast

• According to the provided answer key, the four options are: Upload File, User Criteria, HR Profiles, HR Criteria (options B, D, F, G).

• Upload File: Filter by a file you upload (e.g., CSV) that provides data for bulk case creation.
• User Criteria: Filter by criteria related to users (e.g., role, department) to select target users.
• HR Profiles: Filter by HR profile data associated with users.
• HR Criteria: Filter by HR-related criteria or attributes.

• Correct answer: UBA (A)

• Why: UBA (User Behavior Analytics) analyzes user activity, builds baselines of normal access patterns, and flags deviations (e.g., accessing sensitive projects not related to the user’s role). This insider-threat detection is exactly what the scenario requires.

• Why not the others:

• What the dataflow options mean:

• Why C fits best:

• What the scenario tests: How monitoring paths on a Universal Forwarder (UF) behave when you deploy an app via the Deployment Server.
• How monitoring is configured: Log paths to monitor are defined in inputs.conf (within apps/local on the UF). Splunk on the UF merges inputs from all apps and local config.
• Key point: Deploying a new app that adds a monitor for a log file does not typically remove existing monitors. The UF will continue monitoring any previously configured paths and will add the new ones from the deployed app.

• Correct answer: C) /var/log/maillog and /var/log/messages.

• As the project progresses, you gain information, refine requirements, and lock down baselines (scope, schedule, cost). This reduces overall risk and uncertainty because more is known and addressed.
• At the same time, changes become more costly over time. Later changes can affect already approved baselines, cause rework, impact the critical path, require more coordination, and incur penalties or delays, all of which raise the cost of making changes.

• Quick context: In a three-tier network (Access, Distribution, Core), roles are:

• Why A is correct: The Distribution layer runs Layer 2 and Layer 3 technologies. It often handles VLAN access (L2) and can perform inter-VLAN routing or connect to the core (L3). So it makes sense for the distribution layer to support both L2 and L3.

• Why B is correct: The Core layer is designed to maintain continuous connectivity when devices fail. It is built for high availability and redundancy, providing multiple paths so the network remains reachable even if some components fail.

• Why the others are not correct:

• Answer: C – netstat -an

• netstat -an lists all active connections and listening ports with numeric addresses, without resolving names. This lets you spot DoS patterns such as a flood of connections to a specific port or many connections from a single or few IPs.
• The other options don’t help as directly:

• Windows: netstat -an and look for numerous ESTABLISHED connections to a service or many SYNs to a port.
• Filter by port, e.g., netstat -an | findstr :80, or on Linux: ss -tunap | grep ':80'.

• Netstat provides a snapshot. DoS detection often benefits from continuous monitoring, flow data, and IDS/IPS logs.

• What the scenario asks: You have a Power BI semantic model that pulls from multiple sources (Azure SQL Database and SharePoint Online), must support a single dataset for three reports, keep data current (daily refresh by 7 AM PT), and apply RLS to limit sales data by region. You need to choose a storage mode for the tables.

• Correct choice: Dual storage mode.

• Why Dual fits:

• Key concepts:

• A assigned without var becomes a global variable (in non-strict mode). So after myFunction(), a = 5 is accessible.
• b is declared with var inside the function, so it’s local to that function (function-scoped). It’s not defined in the outer (global) scope.

• console.log(a) prints 5.
• console.log(b) throws a ReferenceError: b is not defined.

• State is stored remotely in Terraform Cloud/Enterprise, not as a local file.
• Terraform runs (plan/apply) can be executed on remote infrastructure, with results visible in the remote workspace UI.
• Supports workspace-based organization, state locking, and concurrency control.
• Useful for team collaboration, centralized governance, and auditable logs.

• In your configuration, you use a backend block like backend "remote" { ... }.
• You still run terraform init to configure/migrate to the remote backend.
• After setup, terraform plan/terraform apply may be executed remotely, and the outputs/logs appear in the Terraform Cloud/Enterprise UI rather than (only) in your local terminal.

• Other backends store state remotely but typically do not provide remote execution; with the remote backend you get remote runs and a collaborative UI, not just remote state storage.

• AWS IoT Greengrass supports local ML inference using models trained in Amazon SageMaker.
• The model and inference code run on the factory’s local server, so processing continues even when internet connectivity is unavailable.
• A Greengrass component can:

• This provides low latency and avoids sending images to AWS for every inference.

• A: Uses AWS-hosted services and therefore depends on internet connectivity. It would fail during an outage.
• C: Snowball Edge is unnecessary for this use case and adds deployment complexity. A local Greengrass server is the intended edge-inference solution.
• D: Amazon Monitron is designed for monitoring industrial equipment health, not analyzing camera images with a custom SageMaker model.

• A. Custom and standard objects have standard fields. True. Both object types include standard fields (for example, CreatedDate, LastModifiedDate, Owner). Custom objects also have the Name field, which is a standard field on all objects.
• B. Standard objects are included with Salesforce. True. Standard objects (like Account, Contact, etc.) come with Salesforce out of the box.
• C. A new standard object can be created. False. You can create custom objects, but you cannot create new standard objects.
• D. Only standard objects support master-detail relationships. False. Master-detail relationships can be used with both standard and custom objects.

• Copilot??????????
• ???????????????
• ??????????
• ?????????????
• ???????????????????

• Prompts and Suggestions: Retained for 28 days
• User Engagement Data: Kept for Two Years

• 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.

• Why: After listing all resources, you need to assign who is responsible for each task. A RACI chart (Responsible, Accountable, Consulted, Informed) documents these roles for every activity.

• What RACI means:

• Why the others aren’t suitable here:

• Quick example: Task "Create user authentication module" — Responsible: Developer, Accountable: Tech Lead, Consulted: Security Architect, Informed: PM.

• IPv6 address types you’ll typically see:

• Unique Local Addresses (ULAs)

• If an address begins with fdXX:… or fc00:…, place it under “Unique Local Address” (ULA).
• If it begins with 2000::/3, place it under “Global Unicast.”
• If it begins with fe80::, place it under “Link-local.”
• If it begins with ff00::, place it under “Multicast.”

• IPv6 address types:

• The block FF00::/12 is a subset of the multicast space, so it represents multicast addresses.
• Therefore, this block forwards packets to a multicast address (to multiple receivers) rather than to a single unicast address.
• The other prefixes are unicast and would deliver to a single interface.

• Correct answer: D (Special category data)

• Why:

• Quick notes:

• A) MTU mismatch: OSPF neighbors must exchange Hello packets with a compatible MTU. If MTUs differ, the neighbor conversation fails and the adjacency cannot become Full.

• B) Incorrect subnet mask: Routers must be in the same IP subnet on the connected link to form a neighbor relationship. A subnet mismatch prevents Hello packets from being considered valid on the same network.

• D) Area ID not the same: Neighbors must share the same Area ID. A mismatch stops adjacency from forming beyond the 2-Way state.

• E) Different metric set on each end of the link: Cost differences do not prevent OSPF adjacency; they affect routing decisions, not neighbor formation.
• F) Router ID not defined: OSPF can auto-choose a router ID; an undefined ID is not a prerequisite for adjacency to reach Full.

• The two correct options are:

• The two incorrect options:

• Why Ishikawa:

• Why not the others:

• Why: A multi-modal embedding model maps both text and images into a shared vector space, enabling cross-modal similarity and retrieval (e.g., find images matching a text query or text matching an image).

• What it does: Enables efficient similarity search using a vector database or nearest-neighbor indexing.

• Why the other options aren’t suitable:

• Quick implementation note: you’d typically pair the embeddings with a vector store (e.g., ANN index) and possibly fine-tune on domain data for better alignment across modalities.

• Security Development Operations (SDO): a security-focused take on DevOps, integrating security into the development pipeline.
• Service Data Object (SDO): a data representation used in some middleware or enterprise apps.
• Software-Defined Overlay (SDO): a networking overlay concept.
• Secure Data Object: a generic data-security term.

• The Act requires:

• It does not, as a public obligation, mandate publishing a detailed report on the training data used.
• Also, the Act does not specify a fixed “six months” retention for system logs in a way that would clearly make A required; logs for traceability are implied, but the six-month period is not a defined, universal requirement.

• Answer: A. Use Amazon SageMaker Serverless Inference to deploy the model.

• Why: SageMaker Serverless Inference hosts the model and serves real-time predictions without you provisioning or managing servers. It scales automatically and handles deployment details.

• Why the others aren’t suitable:

• Quick example flow: web app ? SageMaker Serverless Inference endpoint ? model ? response.

• Question 142: Security governance is MOST associated with which IT infrastructure component? Network / Application / Platform / Process
• Answer: D. Process

• Security governance is about how decisions are made, accountability, and oversight to ensure security objectives align with business goals. These are implemented through governance processes (policy development, risk management, assurance, monitoring, escalation, etc.).
• The other options (Network, Application, Platform) are concrete IT components; governance affects them, but it’s exercised primarily through the processes that govern how those components are managed and controlled.

• In governance-related questions, focus on the decision-making, accountability, and control processes rather than the underlying technology components.

• Options:

• Answer: The correct choice is A (800-61r3).

• NIST SP 800-61 Rev. 3 is the dedicated guide for incident handling/response and reflects the latest guidance on the incident response lifecycle (preparation, detection/analysis, containment, eradication, recovery, post-incident activity).
• 800-61r2 is an older revision of the same incident response guide.
• 800-53 covers security and privacy controls, not a specific incident response standard.
• 800-171 focuses on protecting CUI in non-federal systems and networks, not incident response guidance.

• The correct answer is A: CCTV recordings are not regularly reviewed.

• CCTV is a monitoring/control mechanism. If recordings aren’t reviewed, security events (break-ins, policy violations, equipment tampering) can go undetected and unaddressed.
• Regular review is essential for deterrence, incident response, and forensic evidence. Without it, the entire CCTV control provides little value.

• B (records deleted after one year) affects forensic availability, but it’s a data-retention issue rather than the ongoing effectiveness of the surveillance program.
• C (footage not recorded 24x7) creates gaps in coverage, which is a concern, but regular review of available footage still supports detection; 24x7 coverage is a separate policy/perimeter issue.
• D (cameras not in break rooms) is more about privacy and space planning than core data-center security monitoring.

• To get min/max per column, you need column profiling. That’s provided by enabling the column profile.
• To actually see those min/max values, you should show the column profile in the details pane, and you also need the details pane itself enabled so you can view it.

• Enable column profile (B)
• Show column profile in details pane (C)
• Enable details pane (E)

• Show column value distribution (D) shows distribution shape but isn’t required to get min/max.
• Show column quality details (D) deals with data quality metrics, not min/max.

• Drop violating frames
• Log or alert on the violation
• Shut down or restrict the port

• Load balancer: Distributes application traffic; it does not control switch MAC learning.
• IPS: May detect some network attacks but is not the primary control for switch MAC-table exhaustion.
• NGFW: Protects traffic crossing network boundaries, while MAC flooding occurs at the switching/LAN layer.

Why B works

• AWS KMS keys are Regional resources. A normal customer managed key in Region A cannot be used directly in Region B.
• A customer managed multi-Region KMS key (MRK) has related key replicas in both Regions. The replicas share key material and key ID, although each has a Regional ARN.
• With client-side encryption, the application encrypts the data before storing it in S3. Cross-Region replication then copies already-encrypted objects, so S3 does not need to decrypt and re-encrypt them with a different Regional key.

Why the other choices are weaker

• A and C: SSE-S3 uses Amazon-managed S3 keys, not the required customer managed KMS key.
• D: SSE-KMS can work only if the solution creates and configures the appropriate KMS key or multi-Region key replica in each Region. As written, it says to create a customer managed KMS key in each Region, which implies separate keys—not the same KMS key. Replication also requires additional KMS permissions and destination encryption configuration.

• Correct answer: C. AWS CloudFormation

• Why: CloudFormation lets you define infrastructure as code via templates (JSON or YAML). You can create a single template that describes multiple AWS resources and reuse it to deploy consistent stacks across environments. Templates can be parameterized and stored for repeated deployments. Other options (Marketplace, AMI, OpsWorks) don’t provide reusable deployment templates in the same way.

• Use a field on Account to show the aggregated yearly shipments (a Roll-Up Summary).
• A Master-Detail relationship between Shipment and Account is required for Roll-Up Summary to work (and to ensure access is governed by the parent Account).
• OWD and the relationship type determine what users can see. If you want to restrict shipment-level access while still showing a summary on Account, Private OWD on Shipment plus Master-Detail helps, because access to shipments is controlled by the Account record.

• A (Public OWD on Shipment with Process Builder and a Lookup to Account): Not ideal. Public access would expose shipments broadly, and a lookup cannot support a built-in roll-up summary.
• B (Private OWD on Shipment, Roll-Up Summary, Master-Detail to Account): Best answer. Private shipments, but because of Master-Detail, access and the roll-up are tied to the Account. The Account can show the yearly total via a Roll-Up Summary field.
• C (Controlled by Parent on Shipment, Trigger, Master-Detail): Master-Detail is correct, but using a trigger is unnecessary for a straightforward roll-up; roll-ups are the standard, maintainable solution.
• D (Private OWD on Shipment, Trigger, Lookup to Account): More complex and not as efficient; no automatic roll-up, and lookup breaks the ease of summarizing by Account.

• The requirement: the support plan must allow opening new support requests (technical support incidents).
• Azure’s Basic plan does not include technical support incidents—only billing/subscription support and access to community support.
• Therefore, you must choose a paid plan to open new technical support requests: Developer, Standard, Professional Direct, or Premier.

• Key concept: ALE = SLE × ARO

• SLE = $15,000
• Occurrences: 2 in 3 years ? ARO = 2/3 per year ˜ 0.6667

• ALE = $15,000 × 0.6667 ˜ $10,000 per year

• ALE is the expected annual loss, so it averages the losses over each year.
• This helps prioritize risk treatments by comparing ALE to the cost of controls.

• Reusing features across different models and teams
• Consistent feature definitions
• Managing feature versions and metadata
• Using features for both training and inference

• SageMaker Data Wrangler: Prepares and transforms data; it is not primarily a shared feature repository.
• SageMaker Clarify: Helps detect bias and explain model predictions.
• SageMaker Model Cards: Documents models, including intended use, risks, and evaluation results.

1. Ingestion Lambda receives requests from API Gateway and places messages in an SQS queue.
2. Processing Lambda is triggered by the queue and loads records into Aurora PostgreSQL.

• Decoupling: API request handling is separated from database loading.
• Buffering: SQS absorbs bursts when incoming traffic exceeds the database’s processing capacity.
• Controlled concurrency: Configure the SQS-triggered Lambda’s concurrency and batch size to avoid overwhelming Aurora or exhausting database connections.
• Automatic retries: Messages that fail processing can return to the queue and be retried. A dead-letter queue can capture persistent failures.
• Lower operational effort: The solution remains serverless and avoids managing EC2 servers or Kubernetes clusters.

• A: EC2 and Tomcat add server-management overhead.
• B: DynamoDB/DAX would require changing the database model and does not solve the Aurora ingestion design.
• C: SNS is useful for notifications and fanout, but SQS is better for durable buffering and backpressure in this producer-consumer workflow.

1. Remote access to a Windows PC
2. Full chat and optional game-console functions

• Windows Remote Desktop (RDP) uses TCP 3389. The router needs a port-forwarding rule that sends incoming traffic on that port to the Windows PC’s internal IP address.
• The game console should connect through the wireless AP. Using WPA2-PSK provides the required wireless security while allowing normal online, chat, and optional services.
• The firewall should separate the internal network from the screened subnet. Its screened-subnet-side address is shown as 10.100.0.1.
• Devices must be placed in the correct network locations: the PC behind the router, the console on the wireless AP, and the firewall between the appropriate network segments.

• 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.

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

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

• Sustainability
• Performance efficiency
• Reliability