Skip to content
Tech Interview Prep
Technical interview guide

Cloud-Native Network Security

Security groups, network ACLs, and private networking that control traffic flow within a cloud environment.

Read
45 min
Practice MCQs
25
Interview Q&A
25
Edition
v4
Editorial status
Reviewed

Scope: AWS VPC, Azure networking/Private Link, Google Cloud NGFW, and NIST SP 800-207 guidance current 2026-08-31.

Curated: · Written: · Reviewed:

25 questions

MCQ-1Host A (10.0.1.7, subnet 10.0.1.0/24) must reach Postgres on host B (10.0.2.5, subnet 10.0.2.0/24). A's routing table holds only `10.0.1.0/24 dev eth0 proto kernel scope link src 10.0.1.7`; there is no default route. B's firewall already allows inbound tcp/5432 from 10.0.1.0/24. Running `psql -h 10.0.2.5` on A fails at once with 'Network is unreachable'. Which change fixes it?
MCQ-3A team must block a known-bad external CIDR from reaching every instance in a subnet, while leaving other traffic from that CIDR untouched. Which statement about network ACLs and security groups explains why the block has to be a network ACL rule?
MCQ-4A client at 10.0.1.50:52344 sends a SYN to a server at 10.0.2.20:443. The server's subnet has a stateless network ACL with inbound rules allowing TCP 443 from 0.0.0.0/0 and outbound rules allowing TCP 443 to 0.0.0.0/0. The client's SYN reaches the server, but the client never receives a SYN-ACK. What is the cause?
MCQ-2A web server on an EC2 instance listens on TCP 443. A client at 203.0.113.10 connects from source port 49152. The client's SYN reaches the instance, but the client never receives a SYN-ACK and the connection times out. The instance's security group has one inbound rule (allow TCP 443 from 0.0.0.0/0) and one outbound rule (allow TCP 443 to 0.0.0.0/0). The subnet's network ACL has one inbound rule (allow TCP 443 from 0.0.0.0/0) and one outbound rule (allow TCP 443 to 0.0.0.0/0). What is the cause?
MCQ-20A service runs dual-stack. Over IPv4 it is healthy. Over IPv6, the TLS handshake completes and responses that fit in a single packet under 1,400 bytes arrive, but any response larger than that stalls and times out. The server's LAN MTU is 1,500; the path to the client crosses an IPsec tunnel with a 1,400-byte MTU. The perimeter firewall's IPv6 ruleset mirrors the IPv4 one: TCP 443 permitted, all ICMPv6 denied. What is the cause?
MCQ-13A VPC flow log with a 10-minute aggregation window contains one ACCEPT record: 198.51.100.7 → 10.0.2.15:443, 4,120 packets, 3.9 MB. The web server on 10.0.2.15 logged 9 requests from 198.51.100.7 in that window, all HTTP 200. An analyst drafts: 'The flow log proves the attacker sent 4,120 requests and the server answered every one.' Which correction is right?
MCQ-5An application runs in a subnet with no route to an Internet Gateway or NAT Gateway. It must call DynamoDB and Secrets Manager, and the security policy forbids that traffic from traversing the public internet. Which design satisfies both constraints?
MCQ-7A VPC has private subnets in three AZs, and all three route tables send 0.0.0.0/0 to a single NAT gateway in us-east-1a. A fleet of 300 instances in those subnets each hold 200 long-lived TCP connections to the same partner endpoint, 203.0.113.10:443. After the fleet scales up, new connections to that endpoint fail with ErrorPortAllocation while the established connections keep working and instance CPU and network throughput stay low. The NAT gateway's documented limit is 55,000 simultaneous connections to a single destination. What is the cause and the fix?
MCQ-6A storage account has public network access set to Deny, and an Azure Policy forbids re-enabling it. A private endpoint for the account was created in subnet `pe-subnet` (10.0.1.0/24) of VNet A; its NIC holds private IP 10.0.1.4. A private DNS zone `privatelink.blob.core.windows.net` exists in the subscription with an A record `mystorage -> 10.0.1.4`. From a VM in VNet A, `nslookup mystorage.blob.core.windows.net` returns 20.150.44.10, and `curl https://mystorage.blob.core.windows.net` times out. What is the cause?
MCQ-23A storage account has a private endpoint at 10.0.0.4 in VNet A (10.0.0.0/16) and public network access enabled for a migration window. The privatelink.blob.core.windows.net zone is linked to VNet A. A new VNet B (10.1.0.0/16) is peered to VNet A with traffic allowed both ways, and both VNets use Azure-provided DNS. A client in VNet B reaches the account, but flow logs show the connection going to 20.60.44.12 over the internet gateway, not to 10.0.0.4. What is the cause?
MCQ-16A payments service calls api.fraudvendor.com on 443. Egress is a security-group allowlist of eight IPs. Last week 22% of calls failed with TCP connect timeouts; every failure went to an address outside the eight, and every call to an allowlisted address succeeded. The vendor serves from a CDN that rotates addresses per request and publishes no stable ranges. Which change fixes the failures while keeping egress limited to that vendor?
MCQ-8A payments API runs on an autoscaling group of 20 instances. The database tier accepts traffic only from firewall rules that list each API instance's private address as a /32 entry; a cron job rebuilds that list every 30 minutes from the cloud API. A deploy at 14:01 replaces all 20 instances within 6 minutes. From 14:07 to 14:30, every new instance's connection attempt to the database hangs and times out; at 14:30 they connect normally with no configuration change. The cron job's log shows a successful run at 14:00 that wrote 20 entries, and the database's pool metrics show 12 of 200 connections in use throughout. Which explanation fits this evidence?
MCQ-25A platform team replaces per-pod IP allowlists in its L3/L4 network policy with mesh mTLS between services. A pod is rescheduled and receives a new IP, yet the mesh still authenticates its callers. Which property of mesh mTLS explains that, and what does it not by itself provide?
MCQ-9A payments namespace runs a NetworkPolicy that allows ingress to pods labeled app=payments only from pods labeled app=frontend in the same namespace. No CIDR exceptions exist. An attacker has a pod in that namespace labeled app=scanner and RBAC that permits patch on pods in the namespace. They want to reach the payments pods on port 8443. Which action lets them receive the allowed traffic?
MCQ-10A platform team attaches a hierarchical firewall policy at the organization node with a single rule: priority 1000, direction EGRESS, action deny, destination 10.0.0.0/8. In project payments-prod, an engineer adds a VPC firewall rule allowing egress from the service account to 10.20.4.7/32, but connections to 10.20.4.7:443 still time out. What is the cause and the correct fix?
MCQ-14An engineer adds rule 12 to a 400-rule edge policy: permit tcp/8443 from 10.5.0.0/16 to 10.9.0.0/24, inserted above the existing deny. The change ticket asked for app-tier access only; the app tier is 10.5.1.0/24 and the VPN pool is 10.5.200.0/24. The engineer probes from 10.5.1.10 to 10.9.0.20 on 8443, the connection succeeds, and the change is closed. Which probe would have caught the exposure this one missed?
MCQ-11An external proxy Network Load Balancer has a forwarding rule on 443 with a target SSL proxy and a backend service whose instances listen on 8443; the health check also targets 8443. To lock it down, the team replaces the firewall rules on the backend instances with a single rule allowing only tcp:443 from 0.0.0.0/0. After the change, clients still complete the TCP handshake on 443, but every request is reset before a response, and the instances log no connections from the load balancer. What explains this?
MCQ-12An edge service runs on 12 instances behind an Application Load Balancer in a shared VPC (10.0.0.0/16) that also hosts 40 other teams' workloads. The backend security group allows 10.0.0.0/16 on port 8080. A pentest shows requests reaching the backend on 8080 carrying a client-supplied X-Forwarded-For that the ALB never wrote. Which change closes that direct path?
MCQ-15A platform team runs 60 services across two clouds whose private ranges overlap: both use 10.0.0.0/8. Services call each other directly, and each one skips its authorization check when the caller's source IP falls in 10.0.0.0/8. Audit now requires every request to a service to be attributable to the specific calling workload. Which change meets that requirement?
MCQ-18Hub VNet 10.0.0.0/16 runs an Azure Firewall at 10.0.1.4. Spoke A (10.1.0.0/16) and the newly added Spoke C (10.3.0.0/16) are each peered to the hub with 'allow forwarded traffic' enabled, and each spoke subnet has a UDR 0.0.0.0/0 → 10.0.1.4. The firewall's only network rule allows 10.0.0.0/8 → 10.0.0.0/8 on port 1433. A week after Spoke C was added, flow logs show 10.1.4.9 connecting to 10.3.7.4:1433, which nobody intended. What made that path work?
MCQ-17A thumbnail service on EC2 fetches user-supplied image URLs with a plain HTTP GET and stores the bytes in S3. CloudTrail shows the instance role's temporary credentials being used from 203.0.113.7, an address outside your VPC, and the fetcher's log contains `GET http://169.254.169.254/latest/meta-data/iam/security-credentials/app-role` returning 200. The instance has both IMDSv1 and IMDSv2 enabled and a metadata hop limit of 1. Which single change stops the credential theft while the thumbnail feature keeps working?
MCQ-21A cluster's DNS logs show pods resolving about 3 queries per second to rotating subdomains under a single external domain, each query carrying a base64 label; the rate stays under the resolver's 20 qps per-client limit. The cluster legitimately resolves 12 external FQDNs and nothing else. Which control design stops the tunnel without breaking those 12 dependencies?
MCQ-19A node runs an eBPF CNI in native routing mode with 25,000 endpoints and about 150,000 new flows per second. The team runs a scaling test: per-packet policy latency is measured with 5,000 CIDR rules, then with 25,000. On the iptables CNI they migrated from, the same 5x growth in rules raised per-packet policy latency from about 8 µs to about 40 µs, roughly linear in rule count. Which result should the eBPF run produce?
MCQ-24A retailer runs 900 production workloads on one flat VLAN. In the first incident a compromised web server reached 240 of them in under three minutes. The team then deployed a host agent enforcing default-deny on all east-west traffic, with an allow-list of 38 documented flows. The web server is the source in 6 of those flows and the destination in 9 others. In the second incident the web server is compromised again and its scan of the local /24 returns nothing. How many workloads can the attacker reach from the web server?
MCQ-22A security group rule `0.0.0.0/0 → tcp/5432` was created 11 days ago on an EC2 instance running PostgreSQL 14. The instance has a public IP. The `postgresql.log` is configured with `log_rotation_age = 1d` and `log_rotation_size = 0`, and the instance's EBS volume holds only the last 24 hours of logs. VPC Flow Logs are enabled for the subnet and retained for 14 days. CloudTrail is enabled. The application's database password is stored in AWS Secrets Manager. Which action set is correct?