CAIC Network Security & Protocols 3 — Questions and Answers
Question 1: An AI inference API is exposed via an API gateway. Which rate-limiting strategy BEST prevents model extraction attacks through repeated queries?
- IP-based rate limiting only
- Token bucket with per-API-key quotas and anomaly detection (Correct answer)
- Blocking all requests after business hours
- Requiring CAPTCHA on every request
Correct answer: Token bucket with per-API-key quotas and anomaly detection
Per-API-key token bucket limiting with anomaly detection catches extraction attempts even when attackers rotate IPs while staying within per-IP limits.
Question 2: Which network protocol vulnerability allows an attacker to intercept AI API calls on a shared network by broadcasting fake MAC address associations?
- DNS spoofing
- ARP poisoning (Correct answer)
- BGP hijacking
- ICMP redirect attack
Correct answer: ARP poisoning
ARP poisoning floods the local network with fake ARP replies, associating the attacker's MAC with a legitimate IP to intercept traffic on the LAN.
Question 3: A zero-trust network architecture for AI workloads requires which core principle when a service requests access to a model endpoint?
- Trust all internal network traffic by default
- Verify identity and authorization explicitly for every request (Correct answer)
- Use VPN tunnels between all services
- Whitelist entire subnets for trusted AI nodes
Correct answer: Verify identity and authorization explicitly for every request
Zero trust mandates explicit verification of identity, device health, and authorization for every request regardless of network location.
Question 4: Which TLS configuration weakness allows a downgrade attack that could expose AI API traffic encrypted with weak ciphers?
- HSTS with long max-age
- Supporting TLS 1.0/1.1 alongside TLS 1.3 (Correct answer)
- Using ECDHE key exchange
- Certificate pinning
Correct answer: Supporting TLS 1.0/1.1 alongside TLS 1.3
Keeping legacy TLS versions enabled allows attackers to force negotiation of weaker protocol versions with exploitable vulnerabilities like POODLE or BEAST.
Question 5: An AI company deploys a BGP-announced IP range for its inference endpoints. What attack could cause global traffic to be rerouted to an adversary's network?
- SYN flood
- BGP route hijacking (Correct answer)
- VLAN hopping
- HTTP request smuggling
Correct answer: BGP route hijacking
BGP route hijacking involves announcing more-specific or false routes that cause internet routers to send traffic intended for the victim to the attacker's infrastructure.
Question 6: Which security control detects when an AI service's network traffic deviates significantly from its established baseline behavior?
- Static firewall rules
- Network anomaly detection / behavioral analytics (Correct answer)
- Port knocking
- MAC address filtering
Correct answer: Network anomaly detection / behavioral analytics
Network anomaly detection systems learn normal traffic patterns and alert on deviations such as unusual data volumes, new destinations, or unexpected protocols.
Question 7: HTTP/2 multiplexing introduces which security risk compared to HTTP/1.1 for AI API gateways?
- Lack of TLS support
- Stream-based attacks like HTTP/2 rapid reset that can overwhelm servers (Correct answer)
- Inability to handle JSON payloads
- No support for authentication headers
Correct answer: Stream-based attacks like HTTP/2 rapid reset that can overwhelm servers
HTTP/2 rapid reset attacks exploit stream multiplexing to send and cancel thousands of requests per second, bypassing traditional rate limiting and causing DoS.
CAIC Network Security & Protocols 3
An AI inference API is exposed via an API gateway.
Which rate-limiting strategy BEST prevents model extraction attacks through repeated queries?