MCQS

1. AI Security

An attacker slightly modifies an image of a stop sign so that an AI-based autonomous vehicle classifies it incorrectly. Which attack is this?

A. Data poisoning
B. Evasion attack
C. SQL injection
D. Model extraction

2. Training Data Attack

An attacker inserts thousands of deliberately mislabeled records into an AI model’s training dataset. What is the primary objective?

A. Increase encryption strength
B. Poison the training process
C. Prevent authentication
D. Improve model generalization

3. CIA Triad

A ransomware attack encrypts an organization’s files so legitimate users cannot access them. Which CIA property is primarily affected?

A. Confidentiality
B. Integrity
C. Availability
D. Authentication

4. Adversarial AI

Which statement best describes an adversarial example?

A. A normal input that improves model accuracy
B. An input deliberately designed to cause an AI model to make an incorrect prediction
C. A backup copy of training data
D. A firewall rule generated by AI

5. Prompt Injection

A student asks an AI assistant to summarize a webpage. Hidden instructions on the webpage tell the AI to reveal confidential information from its connected database. What type of attack is this?

A. Prompt injection
B. DDoS
C. ARP spoofing
D. Credential stuffing

6. AI Model Privacy

An attacker repeatedly queries an AI model to determine whether a particular person’s record was included in its training dataset. Which attack is most closely associated with this objective?

A. Membership inference
B. Port scanning
C. DNS poisoning
D. Buffer overflow

7. Model Extraction

An attacker sends thousands of carefully selected queries to an AI API and uses the responses to create a substitute model. This is:

A. Model extraction
B. Data destruction
C. Phishing
D. Sandboxing

8. Backdoor

An AI model behaves normally except when a secret pattern appears in the input, causing a malicious output. What is this most likely?

A. Backdoor attack
B. Brute-force attack
C. DDoS attack
D. Replay attack

9. AI Hallucination

An AI cybersecurity assistant confidently reports a vulnerability that does not actually exist. This is best described as:

A. Encryption
B. Hallucination
C. Authentication
D. Hash collision

10. Secure AI

Which approach provides the strongest protection against AI-generated security decisions being accepted without human verification?

A. Removing all logging
B. Human oversight for high-impact decisions
C. Giving the AI administrator privileges
D. Disabling authenticatio

11. Zero Trust

In a Zero Trust architecture, which principle is most important?

A. Trust users inside the network automatically
B. Never trust, always verify
C. Allow all internal traffic
D. Disable identity verification after login

12. Multi-Factor Authentication

Which combination represents true MFA?

A. Password + username
B. Password + security question
C. Password + fingerprint
D. PIN + password

13. Password Security

Why are passwords normally stored as salted hashes rather than plaintext?

A. To make passwords shorter
B. To allow administrators to read them
C. To make stolen password databases harder to exploit
D. To increase network bandwidth

14. Salt

What is the primary purpose of a cryptographic salt when hashing passwords?

A. Encrypt network traffic
B. Prevent identical passwords from producing identical stored hashes
C. Replace MFA
D. Increase password length automatically

15. Phishing + A

An attacker uses generative AI to create a highly personalized email based on information about an employee’s job and coworkers. What makes this attack particularly dangerous?

A. AI eliminates the need for internet access
B. AI can increase the realism and personalization of social engineering
C. AI automatically encrypts the victim’s files
D. AI prevents credential theft

16. Deepfake Security

A company’s finance officer receives a video call appearing to show the CEO ordering an urgent transfer. What should the employee do first?

A. Immediately transfer the money
B. Verify the request through an independent trusted channel
C. Ask the deepfake for identification
D. Disable the firewall

17. Ransom

Which sequence best represents a typical ransomware impact?

A. Encrypt → deny access → demand payment
B. Authenticate → decrypt → backup
C. Hash → compress → encrypt
D. Scan → patch → recover

18. Dos

Thousands of compromised devices simultaneously send traffic to a website, exhausting its resources. This is:

A. MITM
B. DDoS
C. SQL injection
D. Privilege escalation

19. Botnet

A botnet is best described as:

A. A legitimate cloud database
B. A network of compromised devices controlled by an attacker
C. A type of encryption algorithm
D. A secure password manager

Change block type or style

Move Heading block from position 60 up to position 59

Move Heading block from position 60 down to position 61

Change text alignment

Displays more block tools

20. SQL Injection

A website directly inserts unsanitized user input into a database query. Which vulnerability could result?

A. SQL injection
B. DDoS
C. Shoulder surfing
D. DNSSEC

21. XSS

An attacker injects malicious JavaScript into a webpage that is later executed in another user’s browser. This is:

A. XSS
B. ARP poisoning
C. Ransomware
D. Port knocking

22. Encryption

Which statement about asymmetric cryptography is correct?

A. It uses only one secret key
B. It uses a public/private key pair
C. It cannot support digital signatures
D. It is identical to hashing

23. Digital Signature

A digital signature primarily provides:

A. Confidentiality only
B. Authenticity and integrity
C. Faster internet speed
D. Data compression

24. Hashing

Which property is expected from a secure cryptographic hash?

A. Easy reversal into the original message
B. Predictable collisions
C. Resistance to finding two inputs with the same hash
D. Identical output for every input

25. TLS

HTTPS primarily protects web communication by using:

A. TLS
B. FTP
C. Telnet
D. SMTP only

26. Man-in-the-Middle

An attacker secretly intercepts communication between two parties and potentially alters messages. This is:

A. MITM attack
B. Brute-force attack
C. Backup attack
D. Sandboxing

27. DNS Security

An attacker modifies DNS responses so that a victim is redirected to a malicious website. This is commonly called:

A. DNS cache poisoning
B. Password spraying
C. Model poisoning
D. Data masking

28. Privilege Escalatio

A normal user discovers a vulnerability allowing them to obtain administrator privileges. This is:

A. Privilege escalation
B. Data compression
C. Authentication
D. Encryption

29. Least Privilege

Which security principle states that users and applications should receive only the permissions necessary to perform their tasks?

A. Open access
B. Least privilege
C. Maximum trust
D. Full delegation

30. AI Agent Security

An autonomous AI agent has access to email, databases, cloud storage, and financial systems. What is the greatest security concern?

A. Its interface color
B. Excessive privileges and unintended actions
C. Its response speed
D. Its vocabulary size

Current CISA guidance specifically highlights risks such as privilege escalation, excessive autonomy, accountability gaps, strong identity management, layered defenses, monitoring, and threat modeling for agentic AI. (GovDelivery)

31. Data Poisoning vs Evasio

Which statement correctly distinguishes poisoning from evasion?

A. Poisoning attacks training data; evasion attacks inputs during use
B. Poisoning attacks passwords; evasion attacks firewalls
C. Both only occur after deployment
D. Both only affect encryption

32. AI Supply Chain

An organization downloads a pre-trained AI model from an untrusted repository and integrates it directly into production. What is the biggest concern?

A. Supply-chain compromise
B. Screen resolution
C. Keyboard failure
D. DNS speed

33. Data Leakage

An employee pastes confidential company source code into a public generative AI chatbot. What is the primary security risk?

A. Data exposure
B. DDoS
C. Port scanning
D. Hardware failure

34. Retrieval-Augmented Generation

A chatbot retrieves information from a company’s private documents before generating an answer. Which security control is especially important?

A. Access control on retrieved documents
B. Removing authentication
C. Making all documents public
D. Disabling encryption

35. AI Prompt Leakage

An attacker tries to convince a chatbot to reveal its hidden system instructions. This is commonly associated with:

A. Prompt/system-instruction extraction
B. Disk fragmentation
C. Packet switching
D. Physical security

36. Adversarial Perturbation

An AI image classifier identifies a cat correctly. An attacker adds tiny, carefully designed changes that cause it to classify the image as a dog. These changes are:

A. Adversarial perturbations
B. Encryption keys
C. Digital signatures
D. Authentication tokens

37. Defensive AI

Which AI application can directly assist a Security Operations Center (SOC)?

A. Detecting anomalous network behavior
B. Removing all security logs
C. Disabling endpoint protection
D. Sharing administrator password

38. False Positive

An AI intrusion detection system reports an attack whenever normal employee activity occurs. These incorrect alerts are:

A. False positives
B. True positives
C. True negatives
D. Encryption failures

39. False Negative

An intrusion detection system fails to detect a real cyberattack. This is:

A. False positive
B. False negative
C. True positive
D. Hash collision

40. AI Bias and Security

Why can biased cybersecurity AI be dangerous?

A. It can systematically make inaccurate security decisions for certain cases
B. Bias automatically encrypts data
C. Bias prevents all attacks
D. Bias increases password entropy

41. Explainability

A bank uses AI to identify suspicious transactions. Security analysts need to understand why the AI flagged a transaction. Which property is especially useful?

A. Explainability
B. Compression
C. Latency
D. Bandwidt

42. Model Drift

A fraud-detection AI model becomes less accurate because attackers’ behavior changes significantly over time. This phenomenon is most closely related to:

A. Model/data drift
B. Encryption
C. Hashing
D. Packet fragmentation

43. AI Red Teaming

What is the primary objective of AI red teaming?

A. Identify vulnerabilities and harmful behaviors through controlled adversarial testing
B. Increase employee salaries
C. Remove security controls
D. Guarantee that an AI model can never fail

CISA describes AI red teaming as a structured effort to identify flaws and vulnerabilities in AI systems under controlled conditions. (CISA)

44. Secure-by-Design

Which development practice is most consistent with secure-by-design AI?

A. Add security only after deployment
B. Consider security risks throughout the AI lifecycle
C. Give every component administrator access
D. Ignore training-data provenance

45. AI Data Provenance

Why is tracking the origin of AI training data important?

A. It helps identify whether data is trustworthy and appropriately sourced
B. It makes models smaller
C. It eliminates the need for testing
D. It guarantees perfect prediction

46. Model Exfiltration

An attacker obtains a proprietary AI model’s parameters and architecture without authorization. This is primarily:

A. Model exfiltration
B. Phishing
C. DDoS
D. Password spray

47. AI + Zero Trust

An AI agent requests access to a confidential database. Under Zero Trust, the correct response is:

A. Automatically approve because the AI is inside the corporate network
B. Verify identity, authorization, context, and required permissions
C. Give permanent administrator access
D. Disable database logging

48. Most Dangerous Design

Which AI-security architecture presents the highest inherent risk?

A. AI with read-only access to public data
B. AI with no external connections
C. Autonomous AI with unrestricted access to sensitive systems and the ability to execute actions
D. AI used only for offline educational demonstrations

49. Best Defense Against AI Manipulation

NIST emphasizes that there is no single foolproof defense against adversarial manipulation. Which strategy is therefore most appropriate?

A. Depend entirely on one AI security tool
B. Use layered defenses, monitoring, testing, access controls, and human oversight
C. Disable all cybersecurity controls
D. Assume AI-generated decisions are always correct

NIST notes that AI systems can face attacks across their lifecycle—including manipulated training data, adversarial inputs, and attempts to extract sensitive information—and that current defenses are not foolproof. (NIST Computer Security Resource Center)

50. OLYMPIAD CHALLENGE QUESTION

A company deploys an AI agent that can read emails and access internal databases. An attacker sends an email containing hidden instructions. The AI follows those instructions, retrieves confidential customer information, and attempts to send it externally.

Which combination BEST describes the security problem?

A. Prompt injection + excessive privileges + data exfiltration
B. DDoS + encryption + hashing
C. SQL injection + ransomware + backup
D. DNS poisoning + biometric authentication + compress

Q	Ans	Q	Ans	Q	Ans	Q	Ans	Q	Ans
1	B	11	B	21	A	31	A	41	A
2	B	12	C	22	B	32	A	42	A
3	C	13	C	23	B	33	A	43	A
4	B	14	B	24	C	34	A	44	B
5	A	15	B	25	A	35	A	45	A
6	A	16	B	26	A	36	A	46	A
7	A	17	A	27	A	37	A	47	B
8	A	18	B	28	A	38	A	48	C
9	B	19	B	29	B	39	B	49	B
10	B	20	A	30	B	40	A	50	A

The USA Olympiad