AI Security & Machine Learning Glossary | Giskard

[Live session] Choosing safer LLMs: From LLM benchmarks to your production agents 🚀

July 21, 2026 | 5PM CEST

1

Days

:

00

Hours

:

26

Minutes

:

16

Seconds

Save your spot

🎥 Discover how to red team your AI agents.

Watch demo

.svg)\ \ \ 5,660](https://github.com/Giskard-AI/giskard)

Glossary

Dive into essential terms curated by AI quality, security & compliance experts. Gain clarity in the new language of AI.

Thank you! Your submission has been received!

Oops! Something went wrong while submitting the form.

Filter alphabetical

A

B

C

D

E

F

G

H

I

J

K

L

M

N

O

P

Q

R

S

T

U

V

W

X

Y

Z

Letter

Abductive Logic Programming

ALP uses abduction to hypothesize explanations for observations when knowledge is incomplete.

Learn more

A

Abstract Data Type

ADTs specify what operations a type supports, not how those operations are implemented.

Learn more

A

Accuracy Metric

Accuracy measures the proportion of correct predictions; use with care on imbalanced data.

Learn more

A

ACID Transactions

ACID properties ensure database transactions are reliable and consistent under concurrent load.

Learn more

A

Activation Functions

Activation functions decide whether neurons fire and enable nonlinear learning in neural nets.

Learn more

A

Active Learning in Machine Learning

Active learning selects which data to label next so models improve with fewer annotations.

Learn more

A

Adaptive Gradient Algorithm (AdaGrad)

AdaGrad adjusts learning rates per parameter based on accumulated gradient history.

Learn more

A

Adversarial Machine Learning

Adversarial machine learning covers attacks that fool models and defenses that improve robustness against evasion and poisoning.

Learn more

H

Agent2Agent Protocol

A2A is an open protocol for agents to discover peers, negotiate tasks, and exchange results securely.

Learn more

A

AgentBench Agent Benchmark

AgentBench measures how well language-model agents handle multi-step tasks across diverse environments.

Learn more

A

AgentHarm Safety Benchmark

AgentHarm tests AI agents on multi-step tasks that probe safety boundaries and harm prevention.

Learn more

A

Agentic Chunking

Agentic chunking lets a language model segment content by semantics for smarter retrieval.

Learn more

A

Agentic Orchestration

Agentic orchestration coordinates decisions and actions among humans, agents, and automation layers.

Learn more

A

Agentic RAG

Agentic RAG lets agents actively plan retrieval and tool use beyond a single static RAG pass.

Learn more

A

Agentic Workflow

Agentic workflows use autonomous agents that iterate through planning, tool use, and observation.

Learn more

A

Agent Observability

Agent observability systematically records runtime signals—prompts, plans, tools, and side effects—for reliable AI agents.

Learn more

AI Agent

AI agents autonomously perceive, decide, and act to achieve goals in an environment.

Learn more

A

AI Agent Evaluation

AI agent evaluation is the generate→score→compare→log→improve loop for trustworthy agent releases.

Learn more

A

AI Agent Observability

AI agent observability turns opaque agent decisions into measurable, auditable production signals.

Learn more

A

AI Center of Excellence (AI CoE)

An AI CoE centralizes expertise to guide AI development, governance, and rollout enterprise-wide.

Learn more

A

AI Content Moderation

AI content moderation automatically screens user-generated content for policy violations and safety risks.

Learn more

A

AI Copilots

AI copilots are virtual assistants that boost productivity by automating and guiding tasks in context.

Learn more

A

AI Data Labeling

Data labeling marks inputs and outputs so supervised ML models can learn from examples.

Learn more

A

AI Fairness

AI fairness seeks to prevent biased outcomes and promote equitable treatment in automated decisions.

Learn more

A

AI Model Validation

Model validation confirms an AI system is accurate, reliable, and secure enough for its intended use.

Learn more

A

AI Observability

AI observability provides ongoing visibility into model performance and behavior in production.

Learn more

A

AI Steerability

Steerability means fine-grained control over how AI models behave relative to policies and goals.

Learn more

A

Alignment Metric (NLI)

NLI-based alignment scores whether a text entails, contradicts, or is neutral to a reference.

Learn more

AlpacaEval

AlpacaEval automates evaluation of instruction-following LLMs with toolkits and public leaderboards.

Learn more

A

AlpacaEval Conversation Benchmark

AlpacaEval’s conversation benchmark tests instruction following and response appropriateness.

Learn more

A

ANFIS

ANFIS merges neural nets and fuzzy inference to model complex nonlinear relationships from data.

Learn more

A

Anomaly Detection

Anomaly detection identifies unusual patterns that may signal errors, fraud, or attacks.

Learn more

A

Answer Relevancy Metric

Answer Relevancy scores how well a generated answer addresses the input prompt.

Learn more

Answer Relevancy RAG Metric

In RAG systems, this metric scores how well answers align with the asked question.

Learn more

APPS Coding Benchmark

APPS evaluates how well LLMs solve programming problems with correct, efficient code.

Learn more

A

ARC Reasoning Benchmark

ARC tests whether language models reason through grade-school science questions.

Learn more

A

Argument Correctness Metric

This metric scores logical validity and structure of arguments produced by AI agents.

Learn more

Artificial Neural Network

ANNs learn from data via interconnected units that approximate complex input–output relationships.

Learn more

A

ASCII Smuggling Injection Attack

ASCII smuggling injects hidden instructions via invisible characters that models still parse.

Learn more

Attention in Machine Learning

Attention lets models focus on the most relevant parts of an input when making predictions.

Learn more

A

Attribute

Attributes are the data fields or features that models use to learn patterns and make predictions.

Learn more

A

Auto-Encoders

Autoencoders learn compact representations by encoding data and decoding it back to the original.

Learn more

A

Automated Machine Learning

AutoML streamlines model development by automating search over pipelines and hyperparameters.

Learn more

A

AutoML

AutoML reduces manual ML work by automating model search, tuning, and basic evaluation.

Learn more

A

Autonomous Agents

Autonomous agents perceive context, set goals, and act with little ongoing human control.

Learn more

A

Autoregressive Model

Autoregressive models forecast future points from lagged past values in a sequence.

Learn more

A

Average Precision

Average Precision averages precision across recall levels, often at multiple IoU thresholds for detection.

Learn more

A

AWS Bedrock

AWS Bedrock provides API access to multiple foundation models with enterprise security features.

Learn more

A

AWS Sagemaker

SageMaker provides IDE, training, and deployment tools for cloud machine learning on AWS.

Learn more

A

Backpropagation

Backpropagation calculates gradients through a network so weights can be updated during learning.

Learn more

B

Backpropagation Algorithm

Backpropagation adjusts weights in reverse through the network to minimize prediction error.

Learn more

B

Bagging in Machine Learning

Bagging builds an ensemble from bootstrap samples to stabilize predictions and cut variance.

Learn more

B

Baseline Distribution

Baseline distributions provide the minimum performance bar for evaluating advanced models.

Learn more

B

Baseline Models

Baseline models are simple starting points used to judge whether complex models add real value.

Learn more

B

Batch Normalization

Batch normalization stabilizes training by normalizing interlayer activations within mini-batches.

Learn more

B

Batch Standardization

Batch standardization normalizes outputs between layers to mitigate internal covariate shift.

Learn more

B

Bayes' theorem

Bayes' theorem computes conditional probability from priors and likelihoods.

Learn more

B

Berkeley Function-Calling Leaderboard Domain-Specific Benchmark

Domain benchmark for multi-language LLM function calling, parallel calls, and relevance detection.

Learn more

B

BERT

Bidirectional transformer encoder for contextual NLP representations and fine-tuning.

Learn more

B

Best-of-N Prompt Injection Attack

Adversarial method that samples many prompt variants and keeps any that bypass injection defenses.

Learn more

Bias Metric

Metric for political, gender, and social bias in model outputs used in decisions.

Learn more

Bias Variance Tradeoff

ML tradeoff between underfitting bias and overfitting variance.

Learn more

B

BigBench Reasoning Benchmark

Collaborative benchmark covering logic, math, and language comprehension for LLMs.

Learn more

B

Binary Classification

Supervised task assigning each input to one of two classes.

Learn more

B

Binary Cross Entropy

Log-loss metric for binary probability predictions versus true labels.

Learn more

B

Binomial Distribution

Probability distribution of successes in n identical independent Bernoulli trials.

Learn more

B

Black Box Model

Models whose internal decision mechanics are opaque to users and auditors.

Learn more

B

BLEU

Automatic metric comparing generated text to references via n-gram precision.

Learn more

B

Blue Teaming Capabilities (OWASP Taxonomy)

Defense capabilities: guardrails, runtime firewalls, AI-SPM, and detection that blocks or catches attacks.

Learn more

B

Broken Function Level Authorization Excessive Agency Attack

Probe that tests whether agents execute functions beyond their authorized privilege level.

Learn more

Broken Object Level Authorization Excessive Agency Attack

Probe that checks whether agents fetch unauthorized objects via prompts or tool calls.

Learn more

Calibration Curve

Plot comparing predicted class probabilities to observed frequencies.

Learn more

C

Canonical Schema

Standardized data model enabling consistent exchange across multiple systems.

Learn more

C

Catastrophic Forgetting

Neural-net phenomenon where new learning overwrites previously acquired capabilities.

Learn more

C

Catboost

Gradient boosting toolkit specializing in categorical feature handling.

Learn more

C

Categorical Variables

Features that take values from a finite set of named categories.

Learn more

C

Causal Language Modeling (CLM)

Autoregressive training that predicts the next token from left context only.

Learn more

C

CBRN Harmful Content Attack

Adversarial probes seeking CBRN weapons-related assistance from AI models.

Learn more

Chain-of-Thought

Intermediate step-by-step reasoning used by LLMs; also a security attack surface.

Learn more

C

Chain-of-Thought Evaluation Metric

Metric that scores model outputs using systematic step-by-step reasoning criteria.

Learn more

Chain-of-Thought Prompting

Prompting method that elicits intermediate reasoning steps for complex tasks.

Learn more

C

Chatbot Arena Conversation Benchmark

Human preference benchmark ranking chat models with pairwise votes and Elo scores.

Learn more

C

ChatGLM

Family of bilingual conversational LLMs focused on Chinese–English dialogue.

Learn more

C

CI CD for Machine Learning

Applying continuous integration and delivery to ML training and deployment.

Learn more

C

Citation Framing Injection Attack

Prompt injection that frames malicious asks as academic citations or scholarly references.

Learn more

Classification Threshold

Cutoff on predicted probability used to assign a hard class label.

Learn more

C

Class Imbalance

Training data where one class far outnumbers others, skewing naive metrics.

Learn more

C

Clustering Algorithms

Unsupervised methods that group data points by similarity without labels.

Learn more

C

CodeContests Coding Benchmark

Competitive-programming benchmark for assessing algorithmic code generation by LLMs.

Learn more

C

Code Execution Metric

Metric that evaluates generated code by executing it against tests or expected outputs.

Learn more

Code Interpreter

LLM tool that executes model-generated code, typically inside a sandbox.

Learn more

C

CodeXGLUE Coding Benchmark

Microsoft multi-task benchmark suite for code understanding and generation models.

Learn more

C

CommonsenseQA Reasoning Benchmark

Multiple-choice QA benchmark requiring everyday commonsense reasoning.

Learn more

C

Competitors Brand Damage Attack

Probe that checks if AI can be pushed to endorse competitors or damage brand reputation.

Learn more

Complex Event Processing

Real-time analysis of event streams to detect complex patterns and situations.

Learn more

C

Computer Vision

AI field for interpreting visual data such as images and video.

Learn more

C

Confusion Matrix in Machine Learning

Table summarizing classifier correct and incorrect predictions by class.

Learn more

C

Context Compliance Harmful Content Attack

Multi-turn probe using fabricated conversation history to elicit harmful content.

Learn more

Contextual Precision Metric

RAG metric for whether relevant context ranks above irrelevant chunks.

Learn more

Contextual Recall Metric

RAG metric for whether necessary supporting information was retrieved.

Learn more

Contextual Relevancy Metric

Metric scoring how relevant retrieved or supplied context is to the query.

Learn more

Contextual Relevancy RAG Metric

RAG metric for on-topic retrieved documents before answer generation.

Learn more

Context Window

Maximum token span an LLM can attend to in a single request.

Learn more

C

Continuous Integration Model

CI practice applied to model training, testing, and validation pipelines.

Learn more

C

Continuous Validation

Repeated post-deploy testing of models for quality, safety, and drift.

Learn more

C

Contrastive Learning

Self-supervised method learning by comparing similar vs dissimilar sample pairs.

Learn more

C

Conversational Agent

AI systems that hold multi-turn natural-language conversations with users.

Learn more

C

Convex Optimization

Mathematical optimization where convexity yields tractable global solutions.

Learn more

C

Convolutional Neural Networks

Neural nets with convolutional layers for spatial pattern recognition.

Learn more

C

Copyright Violations Harmful Content Attack

Probe for unauthorized reproduction or distribution of copyrighted materials by AI.

Learn more

Corrective RAG

RAG pattern that grades context and triggers corrective retrieval when needed.

Learn more

C

Correctness Metric

Metric comparing model outputs to expected ground-truth answers.

Learn more

Crescendo Harmful Content Attack

Incremental multi-turn strategy that steers models toward harmful content gradually.

Learn more

Cross-Lingual Language Models

AI models that understand and generate text across multiple languages via shared representations.

Learn more

C

Cross Session Leak Data Privacy Attack

Privacy probe testing whether sensitive data leaks across separate AI sessions.

Learn more

Cross Validation Modeling

A model assessment method that tests on multiple data folds for more reliable generalization estimates.

Learn more

C

Custom LLM Judge Metric

An LLM-as-judge evaluation method with custom criteria and prompts for subjective quality scoring.

Learn more

C

CyberSecEval Harmful Content Attack

A red-team probe using CyberSecEval samples to check if models produce harmful cybersecurity content.

Learn more

DAN Prompt Injection Attack

DAN-style prompt injection uses role-play to push models past safety protocols as an unrestricted assistant.

Learn more

Data Annotation in AI

Labeling data samples so supervised machine learning models can learn from them.

Learn more

D

Data Augmentation

Techniques that synthetically expand training sets with modified copies of existing samples.

Learn more

D

Data Binning

A preprocessing method that maps continuous numbers into discrete bin intervals.

Learn more

D

Data-Centric AI

An approach that improves AI by systematically raising data quality and coverage.

Learn more

D

Data Cleaning

The process of fixing or removing inaccurate, duplicate, or malformed data before analysis.

Learn more

D

Data Decomposition

Separating a time series into trend, seasonal, and residual components.

Learn more

D

Data Flywheel

A loop where usage generates data that improves models and drives further usage.

Learn more

D

Data Granularity

How fine-grained or aggregated the records in a dataset are.

Learn more

D

Data Logging

Automated collection and storage of time-stamped measurements or events.

Learn more

D

Data Mart

A subject-specific subset of a data warehouse for a department or business function.

Learn more

D

Data Purification

Cleaning inaccurate, corrupt, duplicate, or missing data from datasets.

Learn more

D

Data Science Platform

Integrated software for collaborative data prep, modeling, and insight sharing.

Learn more

D

Data Science Techniques

Analytical and modeling methods used to extract actionable insights from data.

Learn more

D

Data Science Tools

Software libraries and platforms used to collect, analyze, and model data.

Learn more

D

Datasets And Machine Learning

Structured collections of examples that train, validate, and test ML models.

Learn more

D

Data Vault

A data warehouse modeling approach built around hubs, links, and satellites.

Learn more

D

Data Versioning

Tracking and labeling dataset states so ML experiments remain reproducible.

Learn more

D

Data Visualizations

Graphical representation of data to reveal trends, anomalies, and patterns.

Learn more

D

Decision Boundary

The region in feature space where a model's predicted class switches.

Learn more

D

Decision Intelligence

A discipline that blends analytics, AI, and decision processes for better outcomes.

Learn more

D

Decision Tree

A model that routes examples through feature-based splits to a leaf prediction.

Learn more

D

Decision Tree In Machine Learning

Supervised models that learn hierarchical splits for classification or regression.

Learn more

D

Decomposed Evaluation Metric

An evaluation approach that scores separate quality dimensions instead of one opaque overall grade.

Learn more

D

Deep Belief Networks

Generative deep models built by stacking restricted Boltzmann machines.

Learn more

D

Deep Learning

Neural network methods with many layers that learn hierarchical representations.

Learn more

D

Deep learning Algorithms

Neural architectures and training methods used to build deep learning systems.

Learn more

D

Deep Q-Network

A reinforcement learning method that uses deep nets to approximate action values.

Learn more

D

Deep Reinforcement Learning

Using deep neural networks as function approximators inside reinforcement learning.

Learn more

D

Deepset Injection Attack

A prompt-injection evaluation that probes whether models follow attacker instructions over system intent.

Learn more

Deep SHAP

A SHAP-based method for explaining predictions of deep neural networks.

Learn more

D

Degradation Model

Statistical or ML models of how performance decays under time or stress.

Learn more

D

Denial of Service Attack

DoS against AI systems via long outputs, loops, or resource-heavy prompts that exhaust capacity or spend.

Learn more

DenseNET

A convolutional architecture with dense skip connections between layers.

Learn more

D

Density Based Clustering

Clustering that finds dense regions of arbitrary shape and marks outliers as noise.

Learn more

D

Diffusion Models

Generative models that create samples by reversing a stepwise noising process.

Learn more

D

Dimensionality Reduction

Techniques that map high-dimensional data into a lower-dimensional representation.

Learn more

D

Direct Preference Optimization

An alignment method that trains LLMs from human preference pairs without a separate reward model.

Learn more

D

Divergent Repetitions Training Data Extraction Attack

An extraction probe that uses divergent repetition patterns to surface memorized or sensitive training content.

Learn more

DoNotAnswer Harmful Content Attack

A harmful-content probe based on DoNotAnswer-style questions that safe AI systems should refuse.

Learn more

Dplyr

An R package for expressive, pipe-friendly tabular data manipulation.

Learn more

D

Drift Monitoring

Tracking distribution or behavior shifts in data and models against a known baseline.

Learn more

D

DROP Reasoning Benchmark

A QA benchmark testing discrete reasoning skills over paragraph-length contexts.

Learn more

D

Early Stopping

A regularization tactic that stops training when held-out metrics plateau or worsen.

Learn more

E

Embedding Projector

Interactive tools for projecting and exploring embedding spaces visually.

Learn more

E

Embedding Similarity Metric

A metric that scores how close two texts are in embedding space for semantic evaluation.

Learn more

Embeddings in Machine Learning

Dense vector representations that encode semantic or structural similarity for ML tasks.

Learn more

E

Embodied Agents

AI agents that perceive and act through a physical or simulated body, not text alone.

Learn more

E

Emotion Detection Metric

A metric that classifies or scores emotional tone in model outputs for UX and safety reviews.

Learn more

Encoding Prompt Injection Attack

Prompt injection that obfuscates harmful instructions via encodings like Base64 or NATO phonetic text.

Learn more

End-to-End Evaluation

Testing the complete AI workflow from user input through tools and final response quality.

Learn more

E

Ensemble Learning

Methods that aggregate multiple models to produce a stronger predictor.

Learn more

E

Enterprise Generative AI

Organization-scale generative AI with governance, security controls, and continuous evaluation.

Learn more

E

Epoch in machine Learning

One full pass of the training set through the learning algorithm.

Learn more

E

Ethical AI

Practices that align AI systems with fairness, transparency, privacy, and human rights.

Learn more

E

ETL Pipeline

Extract–transform–load workflow moving data from sources into analytics systems.

Learn more

E

Evolutionary Algorithms

Population-based optimizers that evolve candidate solutions via selection and variation.

Learn more

E

Exact Match Metric

A precise evaluation method that scores answers via direct string equality to gold labels.

Learn more

Explainable AI (XAI)

AI designed to provide understandable insights into how and why decisions are made.

Learn more

E

Exploratory Data Analysis (EDA)

Foundational analysis that reveals dataset structure, often with visualizations.

Learn more

E

F1-Score Metric

Harmonic mean of precision and recall for evaluating classification and extraction quality.

Learn more

Facial Recognition

Biometric tech that matches faces in images or video to known identities.

Learn more

F

Failure Analysis Machine Learning

Systematic diagnosis of why machine learning models make incorrect predictions.

Learn more

F

Faithfulness Metric

Scores how well an answer stays faithful to retrieved or provided source context.

Learn more

False Positive Rate

Proportion of actual negatives incorrectly classified as positive.

Learn more

F

Feature Engineering

Creating and transforming input variables to improve machine learning models.

Learn more

F

Feature Selection

Choosing the most informative subset of features for a model.

Learn more

F

Feature Vector

An ordered set of numeric features representing a single data example.

Learn more

F

Federated Learning

Collaborative training that keeps raw data local and aggregates model updates.

Learn more

F

Feedback Loop

When model predictions alter future data and can amplify bias or error.

Learn more

F

Few-Shot Evaluation Metric

Assesses model performance when learning from a small number of in-context examples.

Learn more

Few-Shot Learning

ML paradigm where models adapt to new tasks from only a few labeled examples.

Learn more

F

Few-Shot Prompting

Prompting technique that includes a few demonstrations to guide LLM behavior.

Learn more

F

FinBen Domain-Specific Benchmark

A domain benchmark for assessing LLM performance on financial tasks.

Learn more

F

Folium

Python wrapper for creating interactive Leaflet maps from geospatial data.

Learn more

F

F-score

A performance metric combining precision and recall into a single score (often F1).

Learn more

F

GAIA Agent Benchmark

A benchmark of real-world questions for evaluating general AI assistants and agents.

Learn more

G

Gaussian Distribution

The normal bell-curve distribution parameterized by mean and variance.

Learn more

G

Gaussian Mixture Model

Probabilistic model representing data as a mixture of Gaussian components.

Learn more

G

GCG Injection Harmful Content Attack

Greedy Coordinate Gradient (GCG) attacks optimize adversarial suffixes that jailbreak LLM safety filters.

Learn more

Generalist Language Model

An LLM designed for broad tasks across domains rather than one specialized use case.

Learn more

G

Generalized Linear Models

A statistical family linking linear predictors to diverse response distributions.

Learn more

G

Generative Adversarial Networks

Two-network architecture where a generator learns by fooling a discriminator.

Learn more

G

Generative Agents

LLM-based agents that simulate believable behavior via memory, planning, and reflection.

Learn more

G

G-Eval

An LLM-as-judge method that scores outputs via structured evaluation rubrics.

Learn more

G

GOAT Attack Harmful Content Attack

GOAT automates multi-turn adversarial prompting to uncover jailbreaks and harmful-content failures in LLMs.

Learn more

Gradient Boosting

Sequential ensemble method that adds weak learners to correct prior residuals.

Learn more

G

Gradient Descent in Machine Learning

Iterative optimization that follows gradients to minimize training loss.

Learn more

G

Grandma Framing Injection Attack

Emotional framing that wraps harmful asks as innocent family scenarios to jailbreak LLMs.

Learn more

Grid Search

Brute-force search over a specified hyperparameter grid.

Learn more

G

Groundedness Metric

Evaluates whether answer claims are supported by provided or retrieved context.

Learn more

Ground Truth

Authoritative reference data against which model predictions are scored.

Learn more

G

Grouped Query Attention

Transformer attention variant that groups queries to reduce KV cache size.

Learn more

G

GSM8K Math Benchmark

Benchmark of grade-school math word problems for multi-step reasoning.

Learn more

G

Hallucination Index

A composite score for how often an LLM fabricates facts without support.

Learn more

H

Hallucination Metric

An evaluation measure for unsupported or fabricated claims in LLM outputs.

Learn more

Handling Outliers

Methods to detect and treat extreme values that distort analysis or training.

Learn more

H

HarmBench Harmful Content Attack

A benchmark suite for measuring LLM resistance to harmful content generation.

Learn more

Hash tables

Fast key–value maps implemented with hashing and collision resolution.

Learn more

H

Hash Tables

Data structure mapping keys to values via hashing for fast average-case lookup.

Learn more

H

HellaSwag Reasoning Benchmark

A multiple-choice benchmark for commonsense natural language inference.

Learn more

H

Hellinger Distance

A metric for comparing probability distributions, related to Bhattacharyya coefficient.

Learn more

H

Helpfulness Metric

Scores how useful and goal-aligned an LLM response is for the user request.

Learn more

Hijacking Excessive Agency Attack

An attack that manipulates an AI agent into acting beyond its intended permissions or tool scope.

Learn more

Holdout Data

A train/test split portion reserved solely for unbiased model evaluation.

Learn more

H

HumanEval

A benchmark of programming problems scoring LLM-generated code via unit tests.

Learn more

H

HumanEval Coding Benchmark

Functional-correctness benchmark for LLM code synthesis using unit tests.

Learn more

H

Human-in-the-Loop Machine Learning

ML workflows that integrate human judgment for labeling, feedback, or oversight.

Learn more

H

Hyperparameter Optimization

Systematic search for configuration values that maximize validation performance.

Learn more

H

Hyperplane

A flat (n−1)-dimensional subspace used as a linear decision boundary in ML.

Learn more

H

Illegal Activities Harmful Content Attack

Red-team probe for prompts that push AI toward illegal or harmful criminal content.

Learn more

Image Data Augmentation

Artificially expanding image datasets with transforms to improve vision model generalization.

Learn more

I

Image Data Collection

Gathering and organizing images for training computer vision models.

Learn more

I

Imbalanced Data

Datasets where class frequencies are highly uneven, biasing naive classifiers.

Learn more

I

Impersonation Brand Damage Attack

Probe for AI impersonation of people or brands that could damage reputation.

Learn more

In-Context Learning

Adapting an LLM to a task via examples in the prompt rather than fine-tuning.

Learn more

I

Independent and Identically Distributed Data IID

Data where samples are independent and share the same distribution.

Learn more

I

Information Retrieval

Systems that find and rank documents relevant to a user query.

Learn more

I

Instruction Tuning

Fine-tuning LLMs on instruction–response pairs to improve instruction following.

Learn more

I

Intelligent Document Processing IDP

OCR + NLP + ML pipelines that extract structured data from documents.

Learn more

I

Intent Classification Metric

A metric that checks whether text is mapped to the correct intent label.

Learn more

Intersection over Union (IoU)

Overlap ratio between predicted and ground-truth bounding boxes.

Learn more

I

JSON Validation Metric

Automated checks that generated JSON is valid and includes required keys.

Learn more

Juries of Models Metric

Evaluating outputs with multiple LLM judges for more stable scores.

Learn more

J

Keras

A user-friendly high-level API for defining and training neural networks.

Learn more

K

K-Nearest Neighbor KNN

A lazy supervised algorithm that classifies via the K closest labeled examples.

Learn more

K

KNN Models

Non-parametric models that predict from the K nearest training examples.

Learn more

K

Kolmogorov-Smirnov Test

A nonparametric test comparing sample distributions via maximum CDF distance.

Learn more

K

KYC Process

Identity verification processes that deter fraud and meet financial regulations.

Learn more

K

LangChain

Open-source framework for building LLM chains, agents, and tools.

Learn more

L

Language Classification Metric

A metric that detects which natural language a text sample is written in.

Learn more

Large Action Models

Models that map natural-language goals into executable actions in digital environments.

Learn more

L

Learning Rate in Machine Learning

A hyperparameter controlling how much weights update on each training step.

Learn more

L

Learning-to-Rank

Supervised ML that optimizes the order of search or recommendation results.

Learn more

L

LegalBench Domain-Specific Benchmark

A 162-task benchmark covering six types of legal reasoning for AI evaluation.

Learn more

L

Liability Engagement Legal Risk Attack

Probe for AI agents making commitments that create legal or financial liability.

Learn more

LightGBM

A leaf-wise gradient boosting library optimized for speed on tabular data.

Learn more

L

Likert Framing Injection Attack

Jailbreak that frames harmful queries as Likert-scale research surveys.

Learn more

Linear Regression

A statistical method modeling a dependent variable as a linear combination of predictors.

Learn more

L

Link Injection Data Privacy Attack

Probe for AI generating phishing, malware, or other malicious links.

Learn more

Llama

Meta’s open LLM family used for research and many commercial applications.

Learn more

L

LlamaIndex

Framework connecting datasets to LLMs for context-aware RAG apps.

Learn more

L

LLM Agents

Language-model systems that plan, call tools, and act toward user goals.

Learn more

L

LLM Alignment

Steering LLMs so behavior matches human values, intent, and safety policies.

Learn more

L

LLM APIs

HTTP interfaces that let applications call large language models for text tasks.

Learn more

L

LLM-as-a-Service

Cloud-hosted access to large language models via APIs, without operating your own training or inference clusters.

Learn more

L

LLM Benchmarks

Standardized task suites for comparing language model performance.

Learn more

L

LLM Chatbot Evaluation

Systematic testing of chatbot quality, safety, and user experience.

Learn more

L

LLM Cost

Token, infrastructure, and routing costs of operating large language models.

Learn more

L

LLM Debugger

Tools and practices for diagnosing LLM prompt and output failures.

Learn more

L

LLM Deployment

Integrating LLMs into production with performance, safety, and ops controls.

Learn more

L

LLM Distillation

Training a compact model to mimic a larger LLM’s outputs and skills.

Learn more

L

LLM Embeddings

Dense vector representations of text used for search and retrieval.

Learn more

L

LLM Evaluation

Systematic measurement of LLM accuracy, safety, and usefulness.

Learn more

L

LLM Evaluation Framework

Structured protocol for assessing LLM performance, reliability, and ethics.

Learn more

L

LLM Fine-Tuning

Further training a pre-trained LLM on specialized data for niche tasks.

Learn more

L

LLM Gateway

A proxy layer for routing, auth, and policy across LLM providers.

Learn more

L

LLM Guardrails

Runtime rules and filters that keep LLM apps safe, compliant, and secure.

Learn more

L

LLM Hallucinations

When LLMs produce fluent but false or unsupported statements.

Learn more

L

LLM Inference

Using a trained LLM to generate outputs on new, unseen inputs.

Learn more

L

LLM Interpretability

Methods that explain how LLMs turn inputs into outputs.

Learn more

L

LLM Jailbreaking

Bypassing LLM safety restrictions through adversarial prompts and multi-turn attacks.

Learn more

L

LLM Knowledge Base

Curated content used to ground LLM answers in trusted sources.

Learn more

L

LLM Knowledge Graph

Entity-relation graphs used to ground LLM answers with structured facts.

Learn more

L

LLM Leaderboards

Public rankings of language models on shared benchmark suites.

Learn more

L

LLM Observability

Monitoring prompts, traces, and quality signals for LLM applications.

Learn more

L

LLM Ontology

Concept hierarchies that structure domain knowledge for LLM reasoning.

Learn more

L

LLMOps

Practices and tooling for deploying, monitoring, evaluating, and governing large language models in production systems.

Learn more

L

LLM Orchestration

Coordinating models, tools, and memory into multi-step AI workflows.

Learn more

L

LLM Output Parsing

Extracting structured, machine-usable data from free-form LLM responses for safe integration with tools and APIs.

Learn more

L

LLM Overreliance

Excessive dependence on LLM outputs for decisions that still need human judgment, verification, and domain expertise.

Learn more

L

LLM Parameters

Model weights and decoding settings that control how large language models interpret prompts and generate outputs.

Learn more

L

LLM Playground

An interactive environment for experimenting with LLM prompts, models, and parameters before production integration.

Learn more

L

LLM Product Development

Building products that embed large language models into user-facing workflows with clear value, evaluation, and iteration loops.

Learn more

L

LLM Quantization

Reducing the numeric precision of model weights (and sometimes activations) to shrink size and speed inference with minimal quality loss.

Learn more

L

LLM Red Teaming

Adversarial testing of large language models to find jailbreaks, data leaks, bias, and harmful outputs before deployment.

Learn more

L

LLM Sleeper Agents

Language models with latent, trigger-activated behavior—often from fine-tuning or data poisoning—that can hide backdoors until conditions fire.

Learn more

L

LLM Stack Layers

The layered toolchain for LLM apps: data, model, orchestration, serving, evaluation, and interface components.

Learn more

L

LLM Summarization

Using large language models to condense long documents into shorter text while preserving key information.

Learn more

L

LLM Testing

Systematic evaluation of large language models for accuracy, robustness, safety, and reliability before production use.

Learn more

L

LLM Toxicity

Hate, harassment, or abusive content produced by large language models, often learned from unfiltered training data.

Learn more

L

LLM Tracing

Recording prompts, responses, tool calls, latency, and costs across LLM application runs for debugging and observability.

Learn more

L

Locally Interpretable Model-Agnostic Explanations LIME

A model-agnostic method that explains individual predictions with a simple local surrogate model around the instance.

Learn more

L

Logistic Regression

A classification method that models the probability of categorical outcomes using a logistic (sigmoid) function.

Learn more

L

Long Short Term Memory LSTM

A recurrent neural network architecture that retains long-range sequence information via gated memory cells.

Learn more

L

Low Rank Adaptation of Large Language Models

A parameter-efficient fine-tuning method that adapts LLMs with small low-rank matrices instead of full weight updates.

Learn more

L

Machine Learning

A field of AI where algorithms learn patterns from data to make predictions or decisions without being fully hard-coded.

Learn more

M

Machine Learning Algorithm

A computational method that learns patterns from data to predict or decide on unseen examples.

Learn more

M

Machine Learning as a Service (MLaaS)

Cloud services that provide tools and APIs to build, train, and deploy machine learning models without owning the full stack.

Learn more

M

Machine Learning Bias

Systematic skew in ML predictions that unfairly favors or disadvantages groups, often from biased training data or objectives.

Learn more

M

Machine Learning Checkpointing

Periodically saving model and optimizer state during training to enable recovery, comparison, and rollback.

Learn more

M

Machine Learning Deployment

Integrating trained ML models into production applications and APIs so they serve live predictions.

Learn more

M

Machine Learning Inference

Executing a trained model on new data to produce predictions or generations outside the training loop.

Learn more

M

Machine Learning in Software Testing

Using ML to generate, prioritize, or analyze software tests for better coverage and faster defect detection.

Learn more

M

Machine Learning Lifecycle

End-to-end stages of an ML project: problem framing, data, training, validation, deployment, and monitoring.

Learn more

M

Machine Learning Model Accuracy

The proportion of correct predictions; often paired with precision, recall, and slice-aware metrics for decisions.

Learn more

M

Machine Learning Model Evaluation

Assessing ML model effectiveness with held-out data and metrics like accuracy, AUC, F1-score, and calibration.

Learn more

M

Machine Learning Orchestration

Automating and coordinating ML pipeline stages—from data prep to deployment and monitoring—at scale.

Learn more

M

Machine Learning Pipeline

A structured workflow from data preparation through training, evaluation, deployment, and monitoring of ML models.

Learn more

M

Machine Learning Workflows

Ordered steps for delivering ML solutions: data collection, preparation, training, evaluation, and iteration.

Learn more

M

MAP (Mean Average Precision) Metric

A ranking metric that averages precision across queries, widely used for search and retrieval evaluation.

Learn more

M

Math Framing Injection Attack

A jailbreak pattern that hides harmful intents inside mathematical or academic framing to bypass AI safety filters.

Learn more

MATH Math Benchmark

A benchmark of competition-style math problems used to evaluate LLM reasoning from elementary through high-school topics.

Learn more

M

MBPP Coding Benchmark

A Python coding benchmark with hundreds of problems assessing correctness and core programming skill in generated solutions.

Learn more

M

Mean Absolute Error

Average absolute gap between predictions and true values; a common regression accuracy metric.

Learn more

M

Mean Absolute Percentage Error

Average absolute percentage difference between predicted and actual values; common in forecasting evaluation.

Learn more

M

Mean Squared Error (MSE)

Average of squared prediction errors; a common regression loss and accuracy metric.

Learn more

M

Memory-Augmented Neural Networks

Neural networks extended with external memory for longer-term storage and retrieval during complex tasks.

Learn more

M

Meta Learning

A paradigm where algorithms learn how to learn, enabling fast adaptation to new tasks from limited examples.

Learn more

M

METEOR Score

A translation evaluation metric that aligns outputs to references using stems and synonyms, balancing precision and recall.

Learn more

M

Micro-Models

Compact models focused on narrow tasks or data slices within a larger modeling system.

Learn more

M

Misinformation & Disinformation Harmful Content Attack

An adversarial probe that tries to make AI agents generate or amplify false, misleading, or harmful information.

Learn more

Missing Values in Time Series

Absent observations in sequential datasets that require removal, imputation, or specialized models to avoid biased analysis.

Learn more

M

Mixture of Experts

An architecture that routes each input to a small subset of expert subnetworks for scalable model capacity.

Learn more

M

ML Architecture

The blueprint for how data, models, serving, and monitoring components connect in an ML system.

Learn more

M

ML-Based Scoring Metric

Cost-effective text evaluation with pre-trained scorers instead of LLM-as-judge.

Learn more

M

ML Diagnostics

Systematic identification and remediation of performance, bias, and robustness issues in ML models.

Learn more

M

ML Infrastructure

Compute, storage, and tooling foundations for building, training, and deploying machine learning models.

Learn more

M

ML Interpretability

Understanding and describing how ML models make predictions, beyond raw accuracy scores.

Learn more

M

ML Model Cards

Standardized documentation of an ML model’s use, data, metrics, biases, and limitations.

Learn more

M

ML Model Management

Systems and practices for versioning, approving, deploying, and monitoring machine learning models.

Learn more

M

ML Model Validation

Confirming that models meet quantitative and qualitative requirements on intended data before deployment.

Learn more

M

MLOps

Practices that connect ML development and operations for scalable, governed deployments.

Learn more

M

MLOps for Generative AI

CI/CD, monitoring, and governance adapted for generative AI systems and agents.

Learn more

M

MLOps Framework

A structured methodology for automating and governing the ML lifecycle end to end.

Learn more

M

MLOps Monitoring

Tracking live ML performance, drift, and feedback so production models stay trustworthy.

Learn more

M

ML Performance Tracing

Monitoring model predictions and operational metrics across the lifecycle to find bottlenecks and quality drift.

Learn more

M

ML Scalability

Ability of an ML system to process more data or traffic without unacceptable loss of performance or precision.

Learn more

M

ML Stack

The collection of software tools and frameworks used to develop, deploy, and analyze machine learning solutions.

Learn more

M

MMLU benchmark

Massive Multitask Language Understanding: a broad academic benchmark for LLM capability.

Learn more

M

MMLU Reasoning Benchmark

Multiple-choice academic questions used to probe an LLM knowledge and reasoning breadth.

Learn more

M

Model-Based Machine Learning (MBML)

Tailoring probabilistic models to problems via factor graphs and Bayesian inference.

Learn more

M

Model Behavior

How a model predicts and behaves under real inputs, shaped by data, structure, and training.

Learn more

M

Model Calibration

Adjusting predicted probabilities so they match observed outcome frequencies.

Learn more

M

Model Distillation

Compressing a large teacher model into a smaller student that retains most of its behavior.

Learn more

M

Model Drift

Performance loss when live data or conditions diverge from what the model was trained on.

Learn more

M

Model-Driven Architecture

Software development centered on platform-independent models and transformations.

Learn more

M

Model Explainability

Methods that make ML decisions understandable to stakeholders and auditors.

Learn more

M

Model Fairness

Ensuring ML predictions avoid unjustified disparate impact across groups.

Learn more

M

Model Merging

Combining multiple models weights or adapters into one stronger or more specialized model.

Learn more

M

Model Monitoring

Real-time observation of ML performance to detect failures and meet business requirements.

Learn more

M

Model Observability

Inspecting ML inputs, outputs, and internals to debug and optimize production systems.

Learn more

M

Model Parameters

Learned internal coefficients that determine how a model transforms inputs into predictions.

Learn more

M

Model Registry

Centralized versioning and lifecycle management for production-ready ML models.

Learn more

M

Model Retraining

Updating a trained model with new data to improve performance or adapt to change.

Learn more

M

Model Robustness

A model's ability to stay accurate under noise, shifts, and adversarial conditions.

Learn more

M

Model Selection

Choosing the best model among candidates using validation and resampling techniques.

Learn more

M

Model Tuning

Adjusting hyperparameters to optimize a model's learning process and accuracy.

Learn more

M

MRR (Mean Reciprocal Rank) Metric

Ranking metric focused on how soon the first correct answer appears.

Learn more

M

MT-Bench

A multi-turn benchmark for conversational LLM quality and coherence.

Learn more

M

MT-Bench Conversation Benchmark

Benchmark for how well LLMs stay coherent across conversation turns.

Learn more

M

MTEB

Massive Text Embedding Benchmark for comparing embedding model quality.

Learn more

M

Multi-Class Classification

Predicting one class among three or more possible labels for each instance.

Learn more

M

Multilayer Perceptron (MLP)

Feedforward neural net with one or more hidden layers for classification and regression.

Learn more

M

Multilingual LLM

Large language models trained to understand and generate text in many languages.

Learn more

M

MultiMedQA Domain-Specific Benchmark

Medical QA benchmark combining datasets to score factuality, harm, and bias.

Learn more

M

Naive Bayes Model

Fast probabilistic classifier using Bayes theorem with independent-feature assumptions.

Learn more

N

Natural Language Search

Search that understands user intent from free-form language, not just keyword match.

Learn more

N

Natural Language Understanding

AI methods that let machines interpret meaning and intent in human language.

Learn more

N

NDCG (Normalized Discounted Cumulative Gain) Metric

Ranking metric that rewards relevant items higher up the list, normalized to an ideal order.

Learn more

N

Neural Networks

Layered models of artificial neurons used for learning patterns from data.

Learn more

N

Neural Network Tuning

Optimizing neural net hyperparameters like depth, learning rate, and batch size.

Learn more

N

No-code Low-code ML

Platforms that build ML apps and workflows with visual tools instead of heavy coding.

Learn more

N

Noise in Machine Learning

Unwanted data variation that can mislead training and degrade generalization.

Learn more

N

Noisy Image

Images with unwanted brightness or color variation that may need denoising.

Learn more

N

Normalization in Machine Learning

Rescaling features to a common range (e.g., min-max) for stable model training.

Learn more

N

Normalized Discounted Cumulative Gain

Ranking effectiveness metric balancing relevance and position, normalized to ideal order.

Learn more

N

Nvidia NIM

NVIDIA's inference microservices for deploying optimized AI models at scale.

Learn more

N

Observation ML

A single data instance (feature vector) analyzed or predicted by a learning algorithm.

Learn more

O

One-Hot Encoding

Representing categories as binary columns so models can process categorical features.

Learn more

O

Online Machine Learning

Incremental learning from streaming data so models adapt without full batch retrains.

Learn more

O

Open-Source Machine Learning Monitoring

Community tools for supervising live ML models—drift, performance, and alerts.

Learn more

O

Out-of-distribution

Inputs that differ substantially from training data, often causing unreliable predictions.

Learn more

O

Overfitting in Machine Learning

Learning training noise so closely that performance on new data drops.

Learn more

O

OWASP AI Red Teaming Solutions Landscape Taxonomy

OWASP grid of AI red-teaming capabilities mapped to lifecycle stages from Plan and Test through Operate and Govern.

Learn more

O

OWASP ASI01 Agent Goal Hijack

When email, PDFs, or tool output rewrite an agent objective so the workflow chases the wrong goal.

Learn more

A

OWASP ASI02 Tool Misuse and Exploitation

When agents misuse real tools and permissions after injection or misalignment.

Learn more

T

OWASP ASI03 Identity and Privilege Abuse

Delegation chains and inherited credentials that escalate access without re-checking the user.

Learn more

I

OWASP ASI04 Agentic Supply Chain Vulnerabilities

Runtime-loaded MCP servers, prompt packs, and agent personas you did not fully trust.

Learn more

A

OWASP ASI05 Unexpected Code Execution

Coding agents and vibe-coding tools running shell commands nobody reviewed.

Learn more

U

OWASP ASI06 Memory and Context Poisoning

Bad facts that persist across sessions and compound every future task.

Learn more

M

OWASP ASI07 Insecure Inter-Agent Communication

ASI07: weak auth, spoofed peers, and unencrypted A2A/MCP traffic between agents.

Learn more

I

OWASP ASI08 Cascading Failures

ASI08: one poisoned agent output cascading through a multi-agent workflow before humans can stop it.

Learn more

C

OWASP ASI09 Human-Agent Trust Exploitation

ASI09: agents using urgency or authority cues to push humans into unsafe approvals.

Learn more

H

OWASP ASI10 Rogue Agents

ASI10: misaligned or self-persisting agents that keep acting after they should be revoked.

Learn more

R

OWASP LLM01 (2023) Prompt Injection

The 2023/24 OWASP #1 LLM risk: crafted inputs that override developer instructions.

Learn more

P

OWASP LLM01 (2025) Prompt Injection

OWASP #1 LLM risk: direct, indirect, and automated prompt injection that overrides instructions.

Learn more

P

OWASP LLM02 (2023) Insecure Output Handling

2023 OWASP risk: apps trusting LLM output into browsers, shells, or APIs without sanitization.

Learn more

I

OWASP LLM02 (2025) Sensitive Information Disclosure

When LLMs leak PII, credentials, or internal context through chat, tools, or cross-session memory.

Learn more

S

OWASP LLM03 (2023) Training Data Poisoning

2023 OWASP entry for poisoned training or fine-tuning data that embeds backdoors or bias.

Learn more

T

OWASP LLM03 (2025) Supply Chain

LLM supply-chain risk from models, plugins, datasets, and MCP servers you did not build.

Learn more

S

OWASP LLM04 (2023) Model Denial of Service

2023 OWASP risk: resource exhaustion that degrades LLM availability for legitimate users.

Learn more

M

OWASP LLM04 (2025) Data and Model Poisoning

Poisoned training, fine-tuning, RAG, or memory that steers trusted LLM outputs wrong.

Learn more

OWASP LLM05 (2023) Supply Chain Vulnerabilities

2023 OWASP risk from untrusted models, datasets, plugins, and weights in the LLM stack.

Learn more

S

OWASP LLM05 (2025) Improper Output Handling

When trusted LLM text becomes XSS, SQL injection, or shell execution in downstream systems.

Learn more

I

OWASP LLM06 (2023) Sensitive Information Disclosure

2023 OWASP entry for LLMs revealing confidential data in responses or context.

Learn more

S

OWASP LLM06 (2025) Excessive Agency

When an LLM can call APIs, write data, or send messages without enough human checks.

Learn more

E

OWASP LLM07 (2023) Insecure Plugin Design

2023 OWASP risk from poorly designed LLM plugins that escalate privilege or trust input.

Learn more

I

OWASP LLM07 (2025) System Prompt Leakage

When crafted prompts extract system instructions, guardrails, or tool schemas.

Learn more

S

OWASP LLM08 (2023) Excessive Agency

2023 OWASP risk: LLM agents granted more autonomy and tools than safety controls justify.

Learn more

E

OWASP LLM08 (2025) Vector and Embedding Weaknesses

RAG attack surface: poisoned chunks, weak vector DB access, and manipulated similarity.

Learn more

V

OWASP LLM09 (2023) Overreliance

2023 OWASP risk: people and systems trusting fluent LLM answers without verification.

Learn more

O

OWASP LLM09 (2025) Misinformation

Fluent, authoritative LLM falsehoods that drive bad decisions in finance, legal, or medical use.

Learn more

M

OWASP LLM10 (2023) Model Theft

2023 OWASP risk of stealing model weights or approximating behavior through queries.

Learn more

M

OWASP LLM10 (2025) Unbounded Consumption

Runaway tokens, tool loops, and inference spend that drain budget without a hard crash.

Learn more

U

Pandas and NumPy

Pandas and NumPy are core Python libraries for tabular analysis and fast numerical array computation.

Learn more

P

Parameter-Efficient Fine-Tuning

PEFT adapts LLMs by updating a small parameter set (e.g. LoRA) instead of full fine-tuning.

Learn more

P

Parameter-Efficient Fine-Tuning (Prefix-Tuning)

Prefix-tuning learns continuous prefix vectors while freezing most base LLM weights.

Learn more

P

Pascal

PASCAL VOC is a classic object detection and segmentation benchmark with standardized bounding boxes.

Learn more

P

Pattern Matching

Pattern matching finds exact or regex-defined patterns in sequences of data or tokens.

Learn more

P

Pattern Recognition

Pattern recognition identifies regularities in data so systems can classify or detect structures.

Learn more

P

Permutation Importance

Permutation importance estimates feature influence by measuring performance drop after shuffling a feature.

Learn more

P

Phare Safety Benchmark

Phare is a multilingual LLM safety benchmark for probing harmful content and refusal behavior.

Learn more

P

PII Detection Metric

A PII detection metric scores whether model outputs expose personally identifiable information.

Learn more

PII Leak Data Privacy Attack

A PII leak attack probes whether an LLM reveals personally identifiable information it should withhold.

Learn more

Politeness Metric

A politeness metric scores whether model responses stay courteous and appropriately toned.

Learn more

Politics Topic Harmful Content Attack

Safety probe checking whether an LLM refuses disallowed political harm content.

Learn more

Pooling Layers in CNN

Pooling layers downsample CNN feature maps to reduce spatial size and keep salient activations.

Learn more

P

Population Stability Index

PSI measures how much a feature distribution shifted between a reference and current population.

Learn more

P

Positional Encoding

Positional encoding injects sequence-order information into transformer inputs that lack position.

Learn more

P

Precision in Machine Learning

Precision is the fraction of predicted positives that are actually positive.

Learn more

P

Precision Metric

A precision metric scores how many of the systems claimed positives are correct.

Learn more

Precision-Recall Area Under the Curve (PR AUC)

PR AUC summarizes the precision-recall curve, especially useful on imbalanced classification problems.

Learn more

P

Predictive Model Validation

Predictive model validation estimates generalization via holdout sets, cross-validation, or backtests.

Learn more

P

Preprocessing

Preprocessing transforms raw inputs into features models can learn from.

Learn more

P

Pre-trained Transformer

A pre-trained transformer is trained on large corpora then adapted downstream via fine-tuning or prompting.

Learn more

P

Principal Component Analysis (PCA)

PCA projects data onto orthogonal components that capture maximal variance.

Learn more

P

Probabilistic Classification

Probabilistic classification predicts class probabilities, not only hard labels.

Learn more

P

Programmatic Validation Metric

A programmatic validation metric scores outputs with deterministic rules such as schemas or regex.

Learn more

Prompt Alignment Metric

A prompt alignment metric scores how well a response follows user instructions and constraints.

Learn more

Prompt Chaining

Prompt chaining sequences multiple LLM calls so each steps output feeds the next stage.

Learn more

P

Prompt Engineering

Prompt engineering designs instructions and context that steer LLM behavior toward reliable outcomes.

Learn more

P

Prompt Extraction Internal Information Attack

A prompt extraction attack tries to coerce an LLM into revealing its system prompt or hidden instructions.

Learn more

Prompt Injection

Prompt injection occurs when untrusted text overrides developer instructions and steers model behavior.

Learn more

P

Prompt Playground

A prompt playground is an interactive UI for trying prompts, models, and parameters during development.

Learn more

P

Prototype Model

A prototype model is an early proof-of-concept trained to validate feasibility before production hardening.

Learn more

P

Purple Teaming Capabilities (OWASP Taxonomy)

OWASP purple teaming: map red findings to blue controls and measure what actually works.

Learn more

P

PyTorch

PyTorch is an open-source deep learning framework with dynamic graphs and pythonic tensor APIs.

Learn more

P

RAG Architecture

System design that retrieves external context and feeds it into a generative model for grounded answers.

Learn more

R

RAGAS

An evaluation framework for measuring RAG pipeline quality and reliability.

Learn more

R

RAG as a Service

Cloud-managed RAG that packages retrieval, indexing, and generation for applications.

Learn more

R

RAGET Complex Question Hallucination Attack

Uses complex questions derived from your knowledge base to test RAG hallucination resistance.

Learn more

RAGET Simple Question Hallucination Attack

Generates simple factual questions from your knowledge base to evaluate RAG answer accuracy.

Learn more

RAGET Situational Question Hallucination Attack

Tests RAG hallucinations using situational and user-specific question framing drawn from your docs.

Learn more

RAG Evaluation

How to assess RAG systems across retrieval quality and generated answer faithfulness.

Learn more

R

RAG Hallucinations

When retrieval-augmented models still produce false or unsupported claims that sound credible.

Learn more

R

Random Forest

Ensemble of decision trees using bagging and random feature selection.

Learn more

R

Random Initialization

Initializing network weights randomly near zero to break symmetry for gradient descent.

Learn more

R

Readability Assessment Metric

Scores how easy AI-generated text is for a target audience to read.

Learn more

Reasoning Engine

Component that performs structured inference or planning to produce conclusions or actions.

Learn more

R

Recall in Machine Learning

Fraction of true positives identified among all actual positive cases.

Learn more

R

Recall Metric

The share of true positives found among all actual positives.

Learn more

Recall-Oriented Understudy for Gisting Evaluation (ROUGE)

N-gram recall metrics for comparing generated text to reference summaries.

Learn more

R

Receiver Operating Characteristic (ROC) Curve

Plot of TPR vs FPR used to assess binary classifier discrimination.

Learn more

R

Rectified Linear Unit (ReLU)

Activation that zeros negatives and passes positives unchanged.

Learn more

R

Recurrent Neural Networks

Neural nets with temporal state for sequential data.

Learn more

R

Red Teaming Capabilities (OWASP Taxonomy)

OWASP taxonomy red-teaming capabilities: adversarial probes, multi-turn attacks, and continuous offensive testing for AI systems.

Learn more

R

Reference Distribution

Baseline distribution against which samples or model outputs are compared.

Learn more

R

Regex Pattern Matching Metric

Deterministic checks that match or reject outputs using regular expressions.

Learn more

Regression

Supervised learning that estimates continuous outcomes.

Learn more

R

Regression Algorithms

Family of methods that map features to continuous targets.

Learn more

R

Regularization Algorithms

Techniques that penalize complexity to reduce overfitting.

Learn more

R

Regularization in Machine Learning

Constraining model complexity so predictors generalize to new data.

Learn more

R

Reinforcement learning

Learning optimal actions via rewards and penalties over time.

Learn more

R

Reinforcement Learning from AI Feedback (RLAIF)

Aligning models with preference signals produced by AI judges.

Learn more

R

Relevance Metric

Scores how well results or answers match the user's information need.

Learn more

Religion Topic Harmful Content Attack

Adversarial prompts that provoke biased or harmful statements about religion to test fairness.

Learn more

Reproducible AI

Practices that make AI experiments and evaluations repeatable across runs and teams.

Learn more

R

ResNet

CNN architecture with residual connections enabling deeper training.

Learn more

R

Responsible AI

Principles and practices for building AI that is safe, fair, transparent, and accountable.

Learn more

R

Retrieval-augmented Generation (RAG)

LLM pattern that retrieves external knowledge before generating a response.

Learn more

R

Ridge Regression

Linear model with squared-coefficient penalty for stability.

Learn more

R

RMSprop

Optimizer that scales steps by recent gradient magnitude.

Learn more

R

Robotic Process Automation (RPA)

Automating structured business processes with software robots.

Learn more

R

ROC-AUC Metric

Area under the ROC curve; measures ranking quality independent of a single threshold.

Learn more

R

Root-Cause Analysis

Structured investigation to find underlying causes of failures.

Learn more

R

Root Mean Square Error (RMSE)

Square root of mean squared error between predictions and targets.

Learn more

R

Rotating Proxy

Proxy setup that cycles IP addresses across requests.

Learn more

R

SafetyBench Safety Benchmark

A benchmark suite for measuring large language model safety performance.

Learn more

S

Safety Metric

Evaluation score for how safely an LLM refuses or handles harmful requests.

Learn more

Scikit-learn

Python toolkit for traditional ML algorithms and evaluation.

Learn more

S

Script Injection Data Privacy Attack

Tests whether AI responses can include harmful scripts that enable XSS and data privacy breaches.

Learn more

Segmentation in Machine Learning

Grouping data into meaningful segments for modeling or CV.

Learn more

S

Selective Sampling

Choosing informative samples for labeling to reduce annotation cost.

Learn more

S

Self-Consistency Evaluation Metric

Measures agreement across multiple sampled answers or chain-of-thought paths.

Learn more

S

Semantic Router

Routes queries to tools or agents using semantic similarity or classifiers.

Learn more

S

Semi-supervised Learning

Learning from scarce labels combined with unlabeled examples.

Learn more

S

Sensitivity and Specificity in ML

TPR and TNR metrics describing binary detection performance.

Learn more

S

Sentiment Analysis

NLP task that labels text as positive, negative, neutral, or by emotion.

Learn more

S

Sentiment Analysis Metric

Measures accuracy or agreement of sentiment/polarity predictions.

Learn more

Seq2Seq Model

Encoder-decoder architectures transforming one sequence into another.

Learn more

S

Shadow Deployment

Serving a candidate model on production traffic for comparison without user impact.

Learn more

S

Shapley Values

Game-theoretic attributions explaining each feature's contribution.

Learn more

S

Shell Injection Excessive Agency Attack

Probes for shell injection when AI agents can execute system commands - a key excessive-agency risk.

Learn more

Six-Month Moratorium

2023 proposal to pause training of frontier models beyond GPT-4-class systems.

Learn more

S

Sliding Window Attention

Attention pattern restricting each token to a fixed-size local window for efficiency.

Learn more

S

Softmax Function

Normalizes logits into a probability distribution summing to one.

Learn more

S

SQL Injection Excessive Agency Attack

Tests whether AI agents with DB access can be prompted into unauthorized SQL.

Learn more

SSRF Excessive Agency Attack

Tests AI agents for prompt-induced SSRF through browsing or HTTP tools.

Learn more

Stereotypes and Discrimination Harmful Content Attack

Tests whether models resist prompts that push stereotyped or discriminatory outputs.

Learn more

StrategyQA Reasoning Benchmark

Benchmark of yes/no questions needing implicit multi-step reasoning strategies.

Learn more

S

Summarization Metric

Evaluates summary factual alignment and coverage of essential source content.

Learn more

Supervised Machine Learning

ML paradigm that learns a mapping from features to labels using labeled training data.

Learn more

S

Support Vector Machines (SVM)

Supervised models that separate classes with a maximum-margin hyperplane, often with kernels.

Learn more

S

Surrogate Model

Cheap approximate model standing in for an expensive simulation or black-box function.

Learn more

S

SVAMP Math Benchmark

Benchmark of varied arithmetic word problems for testing LLM math reasoning.

Learn more

S

Sycophancy Hallucination Attack

Tests whether models flip facts to match user framing—sycophantic hallucination.

Learn more

Synthetic Data

Artificially generated data used for ML training, testing, and privacy-preserving validation.

Learn more

S

Synthetic Data Generation

Process of fabricating datasets that mimic real data for ML and privacy use cases.

Learn more

S

Tabular Data

Data organized in rows and columns, as in spreadsheets and relational tables.

Learn more

T

TAP Prompt Injection Attack

Automated jailbreak via Tree of Attacks with Pruning (TAP) iterative prompt search.

Learn more

Task Completion Metric

Measures whether an AI agent successfully completes its assigned end-to-end task.

Learn more

TensorFlow

Open-source deep learning library for training and serving neural networks.

Learn more

T

Test Set in Machine Learning

Data subset reserved to evaluate a final model on unseen examples.

Learn more

T

Text Statistics Metric

Measures fundamental text properties such as length, word count, and character count.

Learn more

Tone Metric

Evaluates whether AI-generated text matches the intended tone and communication style.

Learn more

Tool Correctness Metric

Evaluates whether AI agents select and call the correct tools with appropriate arguments.

Learn more

Top-1 Error Rate

Fraction of cases where the top predicted class is incorrect.

Learn more

T

Topic Classification Metric

Scores how correctly text is categorized into predefined topic labels.

Learn more

Toxicity Metric

Scores whether LLM outputs contain offensive or harmful language for moderation and safety gates.

Learn more

Training Serving Skew

Difference between training-time and serving-time data that harms production performance.

Learn more

T

Transfer Learning

Applies pretrained knowledge to new related tasks, often by fine-tuning.

Learn more

T

Transformer Models

Self-attention neural networks that underpin modern language and multimodal models.

Learn more

T

Transformer Neural Network

Attention-based neural network architecture for sequence modeling tasks.

Learn more

T

Transliteration Prompt Injection Attack

Tests whether script/transliteration obfuscation bypasses AI content filters.

Learn more

Tree-Based Models

Algorithms using decision trees for classification and regression tasks.

Learn more

T

Tree of Thoughts

Prompting method that explores branching reasoning paths for harder multi-step tasks.

Learn more

T

TreeSHAP

Explains tree-based model predictions using Shapley value attributions.

Learn more

T

Triplet Loss Function

Metric-learning loss comparing anchor, positive, and negative inputs.

Learn more

T

True Positive Rate

Proportion of actual positives correctly identified by a classifier.

Learn more

T

Trulens

Open feedback-functions framework for evaluating and monitoring LLM applications.

Learn more

T

TruthfulQA Reasoning Benchmark

Assesses LLM ability to reason toward truthful answers amid common misconceptions.

Learn more

T

TruthfulQA Safety Benchmark

Benchmark for LLM truthfulness against common online misconceptions and false beliefs.

Learn more

T

t-SNE

Technique for projecting high-dimensional data into insightful 2D/3D visualizations.

Learn more

T

Type 1 Error

Incorrectly rejecting a true null hypothesis; also known as a false positive.

Learn more

T

Type 2 Error

Failing to reject a false null hypothesis; also known as a false negative.

Learn more

T

Unauthorized Advice Misguidance Attack

Tests whether AI agents refuse out-of-scope regulated advice under pressure.

Learn more

Uncertainty Quantification

Methods that estimate prediction reliability to support safer AI deployment decisions.

Learn more

U

Underfitting in Machine Learning

Model too simple to capture data patterns, weak on training and new data.

Learn more

U

Unsupervised Learning

Models that infer structure from untagged data without predefined targets.

Learn more

U

Validation Set in Machine learning

Data subset used to tune hyperparameters and select models before final testing.

Learn more

V

Variational Autoencoder

Generative model introducing controlled randomness for data encoding and synthesis.

Learn more

V

Vector Databases

Databases optimized for storing and querying high-dimensional embedding vectors.

Learn more

V

Vertical AI Agents

Industry-specialized AI agents designed for domain workflows and constraints.

Learn more

V

VGGNet

Deep convolutional network known for simplicity and depth in image recognition.

Learn more

V

Vicuna Conversation Benchmark

Evaluates conversational AI on response quality, coherence, and helpfulness across turns.

Learn more

V

Vision Language Models

AI models combining computer vision and NLP for multimodal tasks.

Learn more

V

Word Overlap Metric

Evaluates textual similarity via n-gram overlap methods such as BLEU and ROUGE.

Learn more

XGBoost

Optimized gradient boosting library for speed and performance in ML tasks.

Learn more

X

XSTest Harmful Content Attack

Evaluates AI on XSTest ambiguous questions for calibrated harmful-content refusals.

Learn more

YOLO (Object Detection Algorithm)

Fast, accurate algorithm that detects objects in real time in a single pass.

Learn more

Y

Zero-Shot Learning

Models recognize or solve unseen categories using related prior knowledge.

Learn more

Z

Previous Next

No results

Get AI security insights in your inbox

##### No vulnerabilities found?
 We refund the assessment.

Limited-Time Offer: available until the end of September.

Business email*

Company name*

What specific AI use case or risk are you targeting?*

Demo Video V2 from Giskard AI on Vimeo

Playing in picture-in-picture

Play

Settings

Speed

QualityAuto

2x

1.5x

1.25x

1x

0.75x

0.5x

Picture-in-Picture

Fullscreen

Hit ▶️ now for your full Hub demo! Full video in your email.