Domain 1: AI Concepts & Security Fundamentals (10% of exam)

AI vs ML vs Deep Learning

AI is the broad goal of machines mimicking intelligence; ML is a subset where systems learn patterns from data instead of explicit rules; Deep Learning is a subset of ML using multi-layer neural networks.

Supervised learning

Trained on labeled data (input→correct output pairs), such as a spam/not-spam classifier.

Unsupervised learning

Finds patterns/clusters in unlabeled data, for example, grouping similar network traffic without predefined categories.

Reinforcement learning

An agent learns by trial and error via rewards/penalties, like a system improving firewall rule tuning based on feedback.

Evolutionary computation

Optimization inspired by biological evolution (mutation/selection) to iteratively improve a solution.

Neural networks

Layered, weighted nodes modeled on brain neurons; the base architecture for deep learning.

CNN (Convolutional Neural Network)

Specialized for grid-like data (images); detects spatial patterns like edges and shapes.

RNN (Recurrent Neural Network)

Processes sequences (text, time series) by retaining memory of prior inputs.

Transformer

Processes entire sequences at once using attention to weigh relationships between all tokens.

GPU

Parallel processing chip good for matrix math, used in training and inference of AI models.

TPU

Google-designed chip purpose-built for ML tensor operations.

GenAI

Models that generate new content (text, image, audio) rather than just classify or predict.

Multi-modal models

Models that handle more than one data type at once (e.g., text plus image).

Prompt engineering

Crafting input text to reliably obtain the desired model output or behavior.

RAG (Retrieval-Augmented Generation)

Model retrieves relevant external documents at query time and integrates them into its context before generating answers.

LoRA (Low-Rank Adaptation)

Efficient fine-tuning method that trains small added matrices instead of the whole model.

PEFT (Parameter-Efficient Fine-Tuning)

Techniques that adapt a model by updating only a small subset of its parameters.

Agentic AI

Systems that autonomously plan and take multi-step actions instead of responding once.

Session persistence/memory

An agent retaining context/state across interactions rather than starting fresh each time.

NLP (Natural Language Processing)

The field of enabling computers to parse and generate human language.

Embeddings

Numeric vector representations of text/data capturing semantic meaning, with similar meanings as nearby vectors.

Context window

The amount of text (tokens) a model can see and consider at one time during a single interaction.

Token management

Tracking and limiting the units of text (tokens) a model processes, affecting cost, performance, and context limits.

Vector databases

Specialized databases storing embeddings for fast similarity search.

Hallucination

A model generating false or fabricated information confidently.

Non-determinism

Same input can produce different outputs across runs, unlike traditional deterministic software.

Lifecycle degradation / drift

Changes in a model, data, or behavior over time that reduce accuracy versus original training conditions.

Model collapse

Progressive quality degradation of models trained on AI-generated data instead of original human data.

Evaluation metrics/benchmarks

Standardized tests/scores used to compare model performance.

HITL vs automated execution

Human-in-the-Loop requires human reviews before action, while automated execution acts without human review.

Structured data vs unstructured data

Structured data is organized/tabular (like databases), whereas unstructured data is free-form (like text and images).

Data lifecycle stages

Acquisition, extraction, cleaning, transformation, labeling, and eventual use/retention/deletion.

Data quality standards

Standards for data accuracy, completeness, and relevance.

Data provenance

Documented origin/history of a dataset.

Tokenization

Replacing sensitive data with a non-sensitive placeholder or token.

K-anonymity

Ensuring each record is indistinguishable from at least k-1 others in a dataset.

Masking

Obscuring part of sensitive data (e.g., showing only the last four digits).

Synthetic data

Artificially generated data that mimics real data statistically.

Zero Trust in AI

Never implicitly trust any user, agent, or system component; verify every request continuously.

Prompt injection

Malicious input designed to override or manipulate a model's instructions.

Jailbreaking

Techniques used to bypass a model's safety/content restrictions.

Evasion attack

Crafting input to fool a model's classification.

Exfiltration

Extracting sensitive data from a system via AI.

Model inversion

Reconstructing sensitive training data by analyzing model outputs.

Data poisoning

Intentionally corrupting training data to manipulate model behavior.

Threat frameworks

OWASP, NIST, and MITRE frameworks used for AI risk management and security.

LLM guardrails

Automated checks around a model for input/output validation to prevent misuse.

Responsible AI principles

Principles including transparency, explainability, fairness, and bias awareness.

Model verification techniques

Methods to confirm a model behaves as intended before and after deployment.

Isolation techniques

Sandboxing or segmenting AI environments to contain failures or breaches.

AI in business/software workflows

Embedding AI into existing processes, such as coding assistants or customer service bots.

Industry-specific use cases

Examples of AI applications in different industries, such as healthcare and finance.

AI for security operations

Using AI for threat analysis, vulnerability prioritization, and security automation.