Free Certified AI Cybersecurity Officer Practice Questions

Sample Practice Questions

Try these 10 sample questions from the Certified AI Cybersecurity Officer practice bank. Each includes the answer and explanation. Continue with the interactive practice above to review the full question set.

1. 1What makes a machine-learning approach different from encoding all prediction rules by hand?

   1. A.It learns a predictive model from data.Correct answer: A
   2. B.It turns each input into a fixed lookup key.
   3. C.It removes the need to evaluate predictions.
   4. D.It executes only rules explicitly listed by a user.

   Explanation: The official source states, "Machine learning (ML) is a way to train software, called a model, to make predictions or generate content using data.". Training derives the model from data. A lookup or manually listed rules do not by themselves establish learning, and learned predictions still require evaluation.

   Reference: Google — Introduction to Machine Learning — What is Machine Learning?, introductory definition

2. 2A retrieval assistant encounters instructions hidden in a public webpage and follows them instead of the user’s task. Which attack path is illustrated?

   1. A.Direct access to model weights
   2. B.Indirect prompt injection through retrieved contentCorrect answer: B
   3. C.Training-time label corruption
   4. D.Model compression during deployment

   Explanation: The official source states, "Indirect prompt injection attacks occur when adversaries remotely (i.e., without a direct interface) exploit LLM-integrated applications by injecting prompts into data likely to be retrieved.". The malicious instructions arrive through retrieved data. The scenario does not involve weight access, training labels or compression.

   Reference: NIST Generative Artificial Intelligence Profile — section 2.9, Information Security

3. 3A model-development team wants to reduce the amount of identifiable information used or exposed. Which design approach directly supports that privacy goal?

   1. A.Collect extra identifiers for potential future uses.
   2. B.Retain direct identifiers in every published model output.
   3. C.Use suitable de-identification and aggregation techniques.Correct answer: C
   4. D.Make all training records public to improve transparency.

   Explanation: The official source states, "data minimizing methods such as de-identification and aggregation for certain model outputs, can support design for privacy-enhanced AI systems.". These methods can support privacy by minimizing identifying data. Extra collection, persistent identifiers and public records increase exposure rather than implement that goal; effectiveness still needs contextual assessment.

   Reference: NIST Artificial Intelligence Risk Management Framework — section 3.6, Privacy-Enhanced

4. 4An executive committee claims that selecting an AI supplier transfers the organization’s AI risk decisions entirely to that supplier. Which governance principle contradicts this claim?

   1. A.Only model developers can make decisions about AI risk.
   2. B.Executive accountability ends when a supplier is selected.
   3. C.Risk accountability automatically follows ownership of the training data.
   4. D.Executive leadership retains responsibility for development and deployment risk decisions.Correct answer: D

   Explanation: The official source states, "Executive leadership of the organization takes responsibility for decisions about risks associated with AI system development and deployment.". The framework assigns executive responsibility for these risk decisions. Selecting a supplier does not establish automatic transfer based on development or data ownership.

   Reference: NIST Artificial Intelligence Risk Management Framework — GOVERN 2.3, Table 1, printed page 23

5. 5What normally happens during a tabletop exercise as defined in NIST guidance?

   1. A.Participants execute a live failover of production systems.
   2. B.Participants perform a penetration test on the live service.
   3. C.Participants discuss scenario responses, roles and coordination.Correct answer: C
   4. D.Participants deploy and configure replacement infrastructure.

   Explanation: The official source states, "A tabletop exercise is discussion-based only and does not involve deploying equipment or other resources.". A tabletop is a facilitated scenario discussion of roles, coordination and decisions. Live failover, penetration testing and replacement deployment are operational activities rather than the discussion-based exercise.

   Reference: NIST — Guide to Test, Training, and Exercise Programs — section 2, Types of Exercises

6. 6Which structural feature identifies a deep neural network rather than a network with a single hidden layer?

   1. A.Several output labels
   2. B.More than one hidden layerCorrect answer: B
   3. C.A larger training file
   4. D.An external database connection

   Explanation: The official source states, "A neural network containing more than one hidden layer.". Depth here concerns multiple hidden layers. Output-label count, file size and external connections do not establish that architecture.

   Reference: Google — Machine Learning Glossary — deep model

7. 7An attacker deliberately varies prompts to make a model produce content its output controls normally restrict. What is this behavior called?

   1. A.Training-data poisoning
   2. B.Model inversion
   3. C.JailbreakingCorrect answer: C
   4. D.Model extraction

   Explanation: The official source states, "manipulating prompts to circumvent output controls.". Crafted prompts bypassing output controls are jailbreaking. Poisoning alters training data, inversion aims to recover information about training inputs, and extraction attempts to reproduce model behavior; those are different attack objectives.

   Reference: NIST Generative Artificial Intelligence Profile — section 2.3, Dangerous, Violent, or Hateful Content

8. 8In the federated-learning design described by Google, what do participating devices send to the coordinating server?

   1. A.Raw training examples without model improvements
   2. B.Both raw training examples and model improvements
   3. C.Only a list of prediction labels for the local examples
   4. D.Model improvements instead of the training examplesCorrect answer: D

   Explanation: The official source states, "The devices then upload the model improvements (but not the training examples) to the coordinating server". The described federation aggregates model improvements while examples stay on devices. Uploading raw examples, combining them with updates, or sending only prediction labels does not describe that update exchange.

   Reference: Google — Machine Learning Glossary — federated learning

9. 9Why should an AI governance review involve people beyond the team that built the system, including potentially affected groups?

   1. A.Their perspectives help challenge assumptions about context and impacts.Correct answer: A
   2. B.Their participation substitutes for technical testing of the system.
   3. C.Their agreement transfers organizational risk ownership to affected users.
   4. D.Their attendance serves as continuing approval for all later uses.

   Explanation: The official source states, "Gathering such broad perspectives can help organizations proactively prevent negative risks". Broad engagement helps reveal context assumptions and impacts. Participation does not replace technical testing, transfer organizational risk ownership, or approve every later use without reassessment.

   Reference: NIST Artificial Intelligence Risk Management Framework — section 5.2, Map

10. 10A security exercise maps controls for an AI workflow that includes a model, a retrieval source and a third-party API. Which mapping scope follows NIST’s AI risk approach?

    1. A.Map only controls for the model’s parameter file.
    2. B.Exclude external components because suppliers operate them.
    3. C.Treat every component as having the same controls without analysis.
    4. D.Identify and document controls for internal and third-party components.Correct answer: D

    Explanation: The official source states, "Internal risk controls for components of the AI sys- tem, including third-party AI technologies, are identified and documented.". The mapping includes component-specific controls and third-party technologies. Limiting it to weights, excluding suppliers or assuming identical controls leaves the workflow insufficiently assessed.

    Reference: NIST Artificial Intelligence Risk Management Framework — MAP 4.2, Table 2, printed page 27