This is not a replacement for the official study guide. Please read the official study guide first — use this exam prep only to reinforce or practice your understanding.
Free AAISM exam practice questions with answers and explanations, organized by domain and part.
Domain 2: AI Risk Management
Q361 Data scientists copy production records into a notebook platform, and a few lines of code export them outside the organization. Which threat does this BEST illustrate?
- A. Model inversion
- B. Training data leakage
- C. Prompt injection
- D. Adversarial evasion
Answer: B
Training data can leak from notebooks and training platforms that lack exfiltration controls, creating data sprawl. Model inversion reconstructs data from model outputs and evasion manipulates inputs. Prompt injection targets running LLM applications.
Domain 2: AI Risk Management
Q362 Which control is MOST effective against the notebook data exfiltration described in the manual?
- A. Larger training datasets to dilute any copied records over time
- B. Longer minimum password lengths and mandatory rotation for all data scientists and engineers who use the notebook platform
- C. Robust infrastructure and network access controls paired with application-level monitoring
- D. More frequent model retraining to refresh the notebooks regularly
Answer: C
The manual calls for strong infrastructure and networking access controls with application-level monitoring, because a simple line of code can exfiltrate data. Password length or retraining does not stop copying data out. Dataset size is irrelevant.
Domain 2: AI Risk Management
Q363 An attacker modifies the preprocessing routine so that certain records are mislabeled before training. Which entry point for data poisoning does this BEST represent?
- A. Source
- B. Supplier
- C. Preprocessing
- D. In-context training
Answer: C
Poisoned data can be introduced where data preprocessing and cleansing routines are modified. Source means changed where collected, supplier means altered by a broker or in transfer, and in-context training means a poisoned RAG knowledge base.
Domain 2: AI Risk Management
Q364 A poisoned document is inserted into an external knowledge base used for RAG grounding. Which data poisoning entry point does this BEST represent?
- A. Source
- B. In-context training
- C. In transit
- D. Preprocessing
Answer: B
An external knowledge base used in RAG to augment LLM inputs can be poisoned, which is the in-context training entry point. The other points are about collection, transfer or cleansing of training data.
Domain 2: AI Risk Management
Q365 Which two impact categories does data poisoning MOST commonly produce?
- A. Faster inference and lower cost through compression of the training set, with no change to the model's behavior or risk
- B. Higher fairness and lower drift through added diverse samples
- C. Better explainability and lower latency through simpler model design
- D. A backdoor or Trojan to change specific behavior, or sabotage of safety, trust or performance
Answer: D
Poisoning can create a backdoor to alter specific outputs or sabotage the model to undermine safety, trust or performance. The other options describe improvements, not harm.
Domain 2: AI Risk Management
Q366 Which method of poisoning a model BEST involves a third-party pretrained model?
- A. Changing the interface color scheme of the hosted model console
- B. Poisoning the data only at its source of collection before training
- C. Rotating encryption keys on the storage used for model files
- D. Poisoning at the model supplier before further training or fine-tuning
Answer: D
One of the three ways to poison a model is through a model supplier when using a third-party pretrained model for further training or tuning. The other options are not model poisoning methods.
Domain 2: AI Risk Management
Q367 A model is trained with a hidden trigger so that it behaves normally except when a particular pattern appears. Which technical threat is this MOST likely to be?
- A. Model backdoor attack
- B. Model drift
- C. Credential stuffing
- D. Deepfake generation
Answer: A
A model backdoor makes a model behave normally except under a hidden trigger. Drift is a gradual change in data or behavior, deepfakes are synthetic media, and credential stuffing targets authentication.
Domain 2: AI Risk Management
Q368 An attacker repeatedly queries a public API to reproduce a proprietary model's behavior. Which threat does this BEST represent?
- A. Data poisoning of the training set through a supplier feed
- B. Prompt injection through a hidden instruction in user content
- C. Model drift caused by changes in real-world input patterns
- D. Model extraction or stealing via API queries
Answer: D
Systematic queries through an API can replicate a model's functionality, which is model extraction or stealing. Poisoning affects training data, prompt injection manipulates instructions, and drift is a change over time.
Domain 2: AI Risk Management
Q369 An attacker uses model outputs to reconstruct sensitive training records. Which threat does this BEST represent?
- A. Model extraction
- B. Data poisoning
- C. Credential theft
- D. Model inversion
Answer: D
Model inversion recovers information about training data from outputs. Extraction copies the model itself, poisoning corrupts inputs, and credential theft targets identity.
Domain 2: AI Risk Management
Q370 Crafted input that causes a classifier to misread a stop sign is the BEST example of which threat?
- A. Adversarial attack
- B. Data breach
- C. Model drift
- D. Insider misuse
Answer: A
Adversarial attacks use manipulated inputs to cause incorrect output, and in the manual they are tied to DoS-type availability impact. Drift, insider misuse and data breaches do not depend on crafted inputs.