CRAGE Exam Prep Free practice test →

Free CRAGE Practice Questions

10 free, exam-style Certified Responsible AI Governance and Ethics (CRAGE) practice questions with answers and explanations. No signup required. Work through them below, then take the full free CRAGE practice test to study every exam domain.

Question 1

An LLM-powered assistant reads instructions hidden in a webpage retrieved by its search tool, and those instructions hijack its behavior. This is BEST classified as:

  1. Direct prompt injection by the human user at the input layer
  2. Indirect prompt injection via retrieved or referenced content
  3. Standard SQL injection over the application network layer
  4. Differential privacy budget exhaustion across queries over time
Show answer & explanation

Correct answer: B - Indirect prompt injection via retrieved or referenced content

Question 2

Under the EU AI Act, a bank fine-tunes a foundation model and re-markets the result under its own name. The bank is acting as a:

  1. Provider, because it places the modified system on the market
  2. Deployer, because it merely consumes the underlying foundation model
  3. Importer, because the foundation model came from outside the EU
  4. Distributor, because the bank passes the model to its customers
Show answer & explanation

Correct answer: A - Provider, because it places the modified system on the market

Question 3

An LLM agent is wired to call arbitrary external APIs with broad scopes and ends up performing a sensitive action with no human approval. Which OWASP LLM Top 10 risk does this BEST match?

  1. LLM03: Training Data Poisoning across the training corpus
  2. LLM05: Supply Chain Vulnerabilities across components and data
  3. LLM06: Sensitive Information Disclosure from the model output
  4. LLM08: Excessive Agency (over-permissioned LLM agent action)
Show answer & explanation

Correct answer: D - LLM08: Excessive Agency (over-permissioned LLM agent action)

Question 4

A hiring AI selects 50% of male applicants and 35% of female applicants. The disparate impact ratio is:

  1. 1.43, indicating no concern under any U.S. employment guidance
  2. 0.85, well above the threshold and not indicating any concern
  3. Cannot be calculated from the given data with any confidence
  4. 0.70, below the 4/5 threshold and indicating disparate impact
Show answer & explanation

Correct answer: D - 0.70, below the 4/5 threshold and indicating disparate impact

Question 5

An organization claims it 'implements NIST AI RMF' but only writes a one-page policy. Which gap is MOST likely?

  1. It is missing the Map, Measure, and Manage functions in practice
  2. It is over-applying the Manage function while skipping Govern alone
  3. It has implemented Measure too aggressively across the AI program
  4. It has implemented Map too aggressively across the AI program
Show answer & explanation

Correct answer: A - It is missing the Map, Measure, and Manage functions in practice

Question 6

Under GDPR Article 22, an individual subject to a solely automated decision producing legal effects is generally entitled to:

  1. Direct access to the firm's full source code repository on demand
  2. Custody of every record in the firm's complete training dataset ever
  3. Human intervention, the right to express a view, and to contest the decision
  4. Direct access to the firm's GPU clusters used for inference at scale
Show answer & explanation

Correct answer: C - Human intervention, the right to express a view, and to contest the decision

Question 7

Under GDPR, pseudonymized data is:

  1. Outside the scope of GDPR entirely under all settings
  2. Treated identically to fully anonymized non-personal data
  3. Still considered personal data subject to GDPR obligations
  4. Allowed to be transferred globally with no restrictions ever
Show answer & explanation

Correct answer: C - Still considered personal data subject to GDPR obligations

Question 8

An attacker submits subtly modified images to a deployed image classifier, causing misclassification at inference time. This attack is BEST classified as:

  1. Model inversion to reconstruct training data from outputs
  2. Data poisoning attack injected during training set preparation
  3. Membership inference against records used during training
  4. Evasion attack using adversarial examples at inference time
Show answer & explanation

Correct answer: D - Evasion attack using adversarial examples at inference time

Question 9

An organization based outside the EU sells an AI hiring tool to EU-based companies. The EU AI Act:

  1. Does not apply because the organization is outside the EU completely
  2. Applies only to the EU customers and never the non-EU provider firm
  3. Applies extraterritorially to providers placing AI on the EU market
  4. Applies only after each EU member state ratifies the law separately
Show answer & explanation

Correct answer: C - Applies extraterritorially to providers placing AI on the EU market

Question 10

An organization is integrating ISO/IEC 42001 with NIST AI RMF. The MOST appropriate strategy is:

  1. Reject one and use only the other to avoid all conceptual overlap
  2. Map the management system clauses to AI RMF functions and reuse evidence
  3. Run the two as fully separate, disconnected programs with no coordination
  4. Replace ISO/IEC 42001 with the NIST AI RMF for all governance activities
Show answer & explanation

Correct answer: B - Map the management system clauses to AI RMF functions and reuse evidence

What's on the CRAGE exam

The Certified Responsible AI Governance and Ethics (CRAGE) exam is organized into 11 knowledge domains. These free practice questions are drawn from across them so you can see where you're strong and where to study:

  1. AI Foundations and Technology Ecosystem (not publicly weighted)
  2. AI Concerns, Ethical Principles, and Responsible AI (not publicly weighted)
  3. AI Strategy and Planning (not publicly weighted)
  4. AI Governance and Frameworks (not publicly weighted)
  5. AI Regulatory Compliance (not publicly weighted)
  6. AI Risk and Threat Management (not publicly weighted)
  7. Third-Party AI Risk Management and Supply Chain Security (not publicly weighted)
  8. AI Security Architecture and Controls (not publicly weighted)
  9. Building Privacy, Trust, and Safety in AI Systems (not publicly weighted)
  10. AI Incident Response and Business Continuity (not publicly weighted)
  11. AI Assurance, Testing, and Auditing (not publicly weighted)

Ready for the real thing?

Practice hundreds more CRAGE questions with instant scoring, weak-area drills, and full exam simulations.

Start the free practice test See pricing