CompTIA SecAI+ Practice Questions: 2.0 Securing AI Systems

37 free, exam-style CompTIA SecAI+ (CY0-001) practice questions covering 2.0 Securing AI Systems, each with the correct answer and an explanation. Start a timed exam below, or scroll on to read through the questions.

🧠

Mini Exam

15 Questions · 30 minutes

Start Quiz 🚀
🏆

Mock SecAI+ Exam

50 Questions · 100 minutes

Start Quiz 🚀
📅

Daily Quiz

10 Questions · 20 minutes

Start Quiz 🚀
🎯

Standard Practice

25 Questions · 50 minutes

Start Quiz 🚀

2.0 Securing AI Systems practice questions

Answers and explanations are shown. Take the timed exam above to test yourself first.

Q1. Which of the following attacks involves an adversary manipulating the input data used to train a machine learning model to introduce a backdoor or bias?

  • A.Model inversion
  • B.Data poisoning✓ Correct
  • C.Membership inference
  • D.Sponge attack
Explanation: Data poisoning occurs when an attacker injects malicious data into the training set, causing the model to learn incorrect patterns or backdoors. Learn more.

Q2. According to the OWASP Top 10 for LLM, which vulnerability occurs when an attacker manipulates the LLM through crafted inputs to bypass safety guardrails?

  • A.Model Theft
  • B.Prompt Injection✓ Correct
  • C.Supply Chain Vulnerabilities
  • D.Insecure Output Handling
Explanation: Prompt Injection is the direct manipulation of a Large Language Model (LLM) via inputs to make it ignore previous instructions or perform unintended actions. Learn more.

Q3. A security engineer is implementing a gateway control to prevent Denial of Service (DoS) attacks against an internal LLM API. Which control would be MOST effective?

  • A.Prompt templates
  • B.Rate limiting✓ Correct
  • C.Data lineage tracking
  • D.Model quantization
Explanation: Rate limiting (and token limits) restricts the number of requests a user can make in a given timeframe, effectively mitigating DoS attacks aimed at exhausting API resources. Learn more.

Q4. Which attack type aims to reconstruct the training data or extract sensitive features of a specific individual from the model's output?

  • A.Model inversion✓ Correct
  • B.Model skewing
  • C.Prompt injection
  • D.Adversarial examples
Explanation: Model inversion attacks exploit the model's output confidence scores to reconstruct the input data or features used during training. Learn more.

Q5. To secure an AI system's data at rest, which of the following is a standard cryptographic requirement?

  • A.TLS 1.3
  • B.AES-256 encryption✓ Correct
  • C.Data augmentation
  • D.Vector embedding
Explanation: AES-256 is a standard symmetric encryption algorithm used to secure data at rest. TLS is used for data in transit. Learn more.

Q6. An attacker is using an 'indirect prompt injection' technique. How is this attack typically delivered?

  • A.By typing commands directly into the chatbot interface
  • B.By embedding malicious instructions in a website or document that the LLM processes✓ Correct
  • C.By overloading the server with API requests
  • D.By stealing the model weights via a side-channel attack
Explanation: Indirect prompt injection occurs when an LLM processes external content (like a webpage or email) that contains hidden instructions designed to manipulate the LLM. Learn more.

Q7. Which resource would be BEST for a security team to reference when looking for a knowledge base of adversary tactics and techniques specifically targeting AI systems?

  • A.MITRE ATT&CK
  • B.MITRE ATLAS✓ Correct
  • C.CVSS v3.1
  • D.NIST CSF
Explanation: MITRE ATLAS (Adversarial Threat Landscape for Artificial-Intelligence Systems) is specifically designed to categorize attacks against AI/ML systems. Learn more.

Q8. A developer wants to ensure that the LLM only answers questions related to the company's HR policies and refuses all other topics. Which control should be implemented?

  • A.System prompt/System role✓ Correct
  • B.Data anonymization
  • C.Homomorphic encryption
  • D.Federated learning
Explanation: System prompts (or system roles) set the behavior, context, and boundaries for the AI model, instructing it on what it can and cannot discuss. Learn more.

Q9. In the context of AI security, what does 'Jailbreaking' refer to?

  • A.Gaining root access to the server hosting the model
  • B.Crafting prompts to bypass the model's ethical and safety content filters✓ Correct
  • C.Decrypting the model weights
  • D.Escaping a containerized environment
Explanation: Jailbreaking in AI refers to prompt engineering techniques designed to trick the model into bypassing its built-in safety mechanisms and ethical guidelines. Learn more.

Q10. Which data processing technique involves replacing sensitive data identifiers with pseudonyms to protect privacy while maintaining analytical utility?

  • A.Data poisoning
  • B.Data augmentation
  • C.Data masking/Tokenization✓ Correct
  • D.Data pruning
Explanation: Data masking or tokenization replaces sensitive data with non-sensitive substitutes (pseudonyms) to protect PII/PHI. Learn more.

Q11. Which of the following represents a 'Membership Inference Attack'?

  • A.Determine if a specific data record was part of the model's training set✓ Correct
  • B.Stealing the entire model architecture
  • C.Flooding the model with requests to cause a crash
  • D.Injecting bias into the model
Explanation: Membership inference attacks allow an adversary to determine whether a specific individual's data was used to train a machine learning model, posing a privacy risk. Learn more.

Q12. What is the purpose of 'Red Teaming' in the context of LLM security?

  • A.To optimize the model for speed
  • B.To simulate adversarial attacks to identify vulnerabilities and harmful outputs✓ Correct
  • C.To clean the data before training
  • D.To monitor the cost of API tokens
Explanation: AI Red Teaming involves ethical hackers attempting to force the model to generate harmful content, bypass guardrails, or reveal sensitive info to improve security. Learn more.

Q13. Which technique allows an attacker to manipulate an AI system by creating input that is visually indistinguishable to humans but misclassified by the computer?

  • A.Adversarial Example / Input Manipulation✓ Correct
  • B.SQL Injection
  • C.Cross-Site Scripting
  • D.Buffer Overflow
Explanation: Adversarial examples are inputs (like images with added noise) designed to cause a machine learning model to make a mistake, often invisible to the human eye. Learn more.

Q14. Which of the following is a supply chain attack vector specific to AI?

  • A.Compromised Pre-trained Model from a public repository (e.g., Hugging Face)✓ Correct
  • B.Phishing email to the CEO
  • C.Weak Wi-Fi password
  • D.Unpatched Windows Server
Explanation: Attackers can upload malicious models (pickled files) to public hubs like Hugging Face. When developers download and load them, arbitrary code can execute. Learn more.

Q15. Which of the following is a method to mitigate 'Prompt Injection' attacks?

  • A.Increasing the temperature of the model
  • B.Using 'Sandboxing' or strictly separating user data from instructions✓ Correct
  • C.Giving the user root access
  • D.Disabling encryption
Explanation: Separating instructions (System Prompts) from user data and using architectural sandboxing helps prevent user input from overriding system commands. Learn more.

Q16. What is a 'Sponge Attack' in the context of AI?

  • A.An attack that absorbs all the data from a database
  • B.A Denial of Service (DoS) attack designed to maximize energy consumption and latency of the model✓ Correct
  • C.Cleaning the data to remove impurities
  • D.A watermarking technique
Explanation: Sponge attacks target the hardware implementation of deep learning models to soak up system resources (energy and latency), causing DoS. Learn more.

Q17. Which of the following is a method to detect 'Hallucinations' in LLM outputs?

  • A.Checking the response confidence level/score✓ Correct
  • B.Encrypting the prompt
  • C.Using a VPN
  • D.Increasing the model temperature
Explanation: Monitoring the response confidence level (log-probs) can help identify when a model is unsure, which often correlates with hallucinations. Learn more.

Q18. What is the primary security concern regarding 'Excessive Agency' in AI agents?

  • A.The agent costs too much money
  • B.The agent is granted permissions to perform actions (like sending emails or executing code) beyond what is necessary✓ Correct
  • C.The agent responds too slowly
  • D.The agent has too many parameters
Explanation: Excessive Agency occurs when an LLM-based agent is given the ability to interface with other systems (plugins, APIs) with excessive permissions, leading to unintended damaging actions. Learn more.

Q19. An adversary uses a 'Transfer Learning Attack'. What are they doing?

  • A.Moving data from one server to another
  • B.Using a pre-trained model as a starting point to train a surrogate model that mimics a target victim model✓ Correct
  • C.Transferring money via AI banking
  • D.Uploading a virus to the cloud
Explanation: In transfer learning attacks, adversaries use pre-trained models to build a surrogate model that mimics the target, allowing them to craft adversarial examples that work on the target. Learn more.

Q20. Which monitoring technique involves checking the output of an LLM for toxic, biased, or policy-violating content before showing it to the user?

  • A.Content moderation / Output filtering✓ Correct
  • B.Input validation
  • C.Rate limiting
  • D.Encryption
Explanation: Output filtering or content moderation (often done by a separate smaller model) ensures that the generated response adheres to safety policies. Learn more.

Q21. Which organization publishes the 'AI Risk Repository' mentioned in the exam objectives?

  • A.MIT (Massachusetts Institute of Technology)✓ Correct
  • B.NASA
  • C.CompTIA
  • D.IEEE
Explanation: The MIT AI Risk Repository is a comprehensive database of AI risks referenced in the SecAI+ objectives. Learn more.

Q22. Which attack hides malicious instructions inside user-supplied content so an LLM ignores the system intent?

  • A.Prompt injection✓ Correct
  • B.SQL indexing
  • C.RAID rebuild
  • D.DNS recursion
Explanation: Prompt injection attempts to override or manipulate model behavior through crafted input. Learn more.

Q23. Which control helps prevent an AI assistant from returning secrets from connected tools?

  • A.Tool permission scoping and output filtering✓ Correct
  • B.Longer response temperature
  • C.Unrestricted file access
  • D.Disabling audit logs
Explanation: Least-privilege tool access and filtering reduce the chance that sensitive tool outputs are exposed. Learn more.

Q24. Which technique tests an AI system by intentionally crafting inputs to cause unsafe or incorrect behavior?

  • A.Adversarial testing✓ Correct
  • B.Disk formatting
  • C.MAC learning
  • D.Printer calibration
Explanation: Adversarial testing evaluates how models behave under malicious, ambiguous, or edge-case inputs. Learn more.

Q25. Which AI security risk involves sensitive examples being exposed from training data?

  • A.Training data leakage✓ Correct
  • B.VLAN pruning
  • C.Drive fragmentation
  • D.NTP skew only
Explanation: Training data leakage can reveal sensitive information memorized or inferred from model training data. Learn more.

Q26. Which control can reduce harmful LLM output before it reaches a user?

  • A.Content safety classifier or guardrail✓ Correct
  • B.Open relay
  • C.Unfiltered tool execution
  • D.Plaintext logging of secrets
Explanation: Guardrails and classifiers can detect and block policy-violating or unsafe outputs. Learn more.

Q27. Which attack poisons examples so a model learns attacker-chosen behavior?

  • A.Data poisoning✓ Correct
  • B.Port mirroring
  • C.ARP aging
  • D.Normal validation
Explanation: Data poisoning manipulates training data to degrade performance or implant unwanted behavior. Learn more.

Q28. Which security control limits damage if an AI tool is tricked into calling an external API?

  • A.Least-privilege API tokens✓ Correct
  • B.Permanent root tokens
  • C.Unbounded spending limits
  • D.Disabled monitoring
Explanation: Scoped tokens restrict what a tool can do even if a prompt injection or workflow error occurs. Learn more.

Q29. Which monitoring signal may indicate an LLM application is under prompt injection attack?

  • A.Repeated requests to reveal hidden instructions or ignore prior rules✓ Correct
  • B.Normal TLS certificate renewal
  • C.Scheduled backups
  • D.Expected DNS cache hits
Explanation: Prompt injection attempts often include language that asks the model to ignore instructions, reveal secrets, or misuse tools. Learn more.

Q30. Which control limits an LLM agent so it can read tickets but cannot delete them?

  • A.Tool-level least privilege✓ Correct
  • B.Higher temperature
  • C.Longer context only
  • D.Unfiltered chain output
Explanation: Tool-level least privilege restricts actions available to the agent and limits damage from mistakes or prompt injection. Learn more.

Q31. Which model attack attempts to infer whether a specific record was part of training data?

  • A.Membership inference✓ Correct
  • B.ARP spoofing
  • C.Port knocking
  • D.RAID striping
Explanation: Membership inference attacks try to determine whether a particular example was used in model training. Learn more.

Q32. Which risk can occur if retrieved documents contain attacker-written instructions?

  • A.Indirect prompt injection✓ Correct
  • B.CPU overheating
  • C.DHCP exhaustion only
  • D.Screen burn-in
Explanation: Indirect prompt injection hides malicious instructions in content that the model later retrieves or summarizes. Learn more.

Q33. Which control reduces the chance an AI assistant exposes another tenant data?

  • A.Tenant isolation controls✓ Correct
  • B.Shared global memory
  • C.Public logs
  • D.Single API key for all tenants
Explanation: Tenant isolation prevents data and permissions from crossing customer or organizational boundaries. Learn more.

Q34. Which term describes intentionally crafted input that causes a model to misclassify?

  • A.Adversarial example✓ Correct
  • B.DNS record
  • C.Patch cable
  • D.Storage snapshot
Explanation: Adversarial examples are inputs designed to cause incorrect model outputs. Learn more.

Q35. Which AI system risk is reduced by human approval before high-impact actions execute?

  • A.Autonomous action causing harm✓ Correct
  • B.Cable attenuation
  • C.Keyboard ghosting
  • D.NTP drift
Explanation: Human-in-the-loop approval can prevent unsafe automated actions in sensitive workflows. Learn more.

Q36. Which AI attack attempts to reconstruct parts of training data from model outputs?

  • A.Model inversion✓ Correct
  • B.MAC flooding
  • C.DNS tunneling only
  • D.VLAN hopping
Explanation: Model inversion attacks attempt to infer sensitive training characteristics or examples from model behavior. Learn more.

Q37. Which control helps stop an AI assistant from making unsupported factual claims?

  • A.Require citations to retrieved trusted sources✓ Correct
  • B.Increase randomness only
  • C.Disable input validation
  • D.Remove all context
Explanation: Requiring source-grounded answers and citations helps users verify outputs and reduces unsupported claims. Learn more.

More CompTIA SecAI+ practice topics

Keep studying CompTIA SecAI+

Full CompTIA SecAI+ practice exam · CompTIA SecAI+ study guide · All CompTIA practice questions

Last updated