10 exam-style questions with answers and explanations, straight from our 1,051-question bank. Tap an answer to check yourself. When you're ready, take the scored version in the free practice test.
The CSPAI exam has 50 questions and runs 1 hour.
These 10 free CSPAI questions are organized by exam domain, so you can see how each part of the Certified Security Professional in Artificial Intelligence blueprint is tested. Reveal the answer and explanation under each question.
During pretraining, an LLM learns by predicting the next token in a sequence drawn from a large web corpus that was never manually annotated. Which learning paradigm does this BEST describe?
Correct answer: B - Self-supervised learning, because the labels come from the data itself
A CISO argues that replacing a hosted API model with a self-hosted open-weight model will straightforwardly improve the organisation's security posture. Which statement is the MOST accurate response?
Correct answer: D - Incorrect, because it removes third-party egress but shifts model and infrastructure risk in-house
Why must a transformer add positional encoding to its input embeddings?
Correct answer: A - Self-attention is order-agnostic and would otherwise treat the input as an unordered set
In scaled dot-product attention, the product of the query and key matrices is divided by the square root of the key dimension. What is the PRIMARY purpose of this scaling?
Correct answer: C - To keep the softmax out of saturation, where gradients vanish
A decoder-only model must not condition its prediction for position t on any token after position t. Which mechanism enforces this constraint?
Correct answer: A - Causal masking of the attention scores
In the standard LLM development pipeline, which phase immediately follows pretraining?
Correct answer: D - Supervised fine-tuning on instruction-response pairs
A team fine-tunes a safety-aligned commercial model on a benign, proprietary dataset of customer-support transcripts. No adversarial or harmful content is present in the dataset. What is the MOST significant security concern?
Correct answer: A - The fine-tuning process may degrade the model's existing safety alignment
Which statement correctly describes how LoRA adapts a pretrained model to a new domain?
Correct answer: C - It freezes the base weights and trains small low-rank adapter matrices alongside them
An internal RAG assistant indexes a corporate wiki. An attacker with wiki edit rights embeds hidden white-on-white text instructing the assistant to append confidential figures to every summary. The assistant later retrieves that page and complies. How is this attack BEST classified?
Correct answer: C - Indirect prompt injection, delivered through retrieved content at inference time
A multi-tenant RAG platform must guarantee that Tenant A can never receive content sourced from Tenant B's documents. Where should this authorisation be enforced?
Correct answer: B - At retrieval time, so that only documents the tenant is entitled to ever enter the context
The CSPAI exam also covers these domains. Drill them in the full free practice test:
The full bank has 1,041 more CSPAI questions with explanations.
Continue in the free practice test →
View plans