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
🎥 Discover how to red team your AI agents.
.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.
A
Abstract Data Type
ADTs specify what operations a type supports, not how those operations are implemented.
A
Accuracy Metric
Accuracy measures the proportion of correct predictions; use with care on imbalanced data.
A
ACID Transactions
ACID properties ensure database transactions are reliable and consistent under concurrent load.
A
Activation Functions
Activation functions decide whether neurons fire and enable nonlinear learning in neural nets.
A
Active Learning in Machine Learning
Active learning selects which data to label next so models improve with fewer annotations.
A
Adaptive Gradient Algorithm (AdaGrad)
AdaGrad adjusts learning rates per parameter based on accumulated gradient history.
A
Adversarial Machine Learning
Adversarial machine learning covers attacks that fool models and defenses that improve robustness against evasion and poisoning.
H
Agent2Agent Protocol
A2A is an open protocol for agents to discover peers, negotiate tasks, and exchange results securely.
A
AgentBench Agent Benchmark
AgentBench measures how well language-model agents handle multi-step tasks across diverse environments.
A
AgentHarm Safety Benchmark
AgentHarm tests AI agents on multi-step tasks that probe safety boundaries and harm prevention.
A
Agentic Chunking
Agentic chunking lets a language model segment content by semantics for smarter retrieval.
A
Agentic Orchestration
Agentic orchestration coordinates decisions and actions among humans, agents, and automation layers.
A
Agentic RAG
Agentic RAG lets agents actively plan retrieval and tool use beyond a single static RAG pass.
A
Agentic Workflow
Agentic workflows use autonomous agents that iterate through planning, tool use, and observation.
A
Agent Observability
Agent observability systematically records runtime signals—prompts, plans, tools, and side effects—for reliable AI agents.
AI Agent
AI agents autonomously perceive, decide, and act to achieve goals in an environment.
A
AI Agent Evaluation
AI agent evaluation is the generate→score→compare→log→improve loop for trustworthy agent releases.
A
AI Agent Observability
AI agent observability turns opaque agent decisions into measurable, auditable production signals.
A
AI Center of Excellence (AI CoE)
An AI CoE centralizes expertise to guide AI development, governance, and rollout enterprise-wide.
A
AI Content Moderation
AI content moderation automatically screens user-generated content for policy violations and safety risks.
A
AI Copilots
AI copilots are virtual assistants that boost productivity by automating and guiding tasks in context.
A
AI Data Labeling
Data labeling marks inputs and outputs so supervised ML models can learn from examples.
A
AI Fairness
AI fairness seeks to prevent biased outcomes and promote equitable treatment in automated decisions.
A
AI Model Validation
Model validation confirms an AI system is accurate, reliable, and secure enough for its intended use.
A
AI Observability
AI observability provides ongoing visibility into model performance and behavior in production.
A
AI Steerability
Steerability means fine-grained control over how AI models behave relative to policies and goals.
A
Alignment Metric (NLI)
NLI-based alignment scores whether a text entails, contradicts, or is neutral to a reference.
AlpacaEval
AlpacaEval automates evaluation of instruction-following LLMs with toolkits and public leaderboards.
A
AlpacaEval Conversation Benchmark
AlpacaEval’s conversation benchmark tests instruction following and response appropriateness.
A
ANFIS
ANFIS merges neural nets and fuzzy inference to model complex nonlinear relationships from data.
A
Anomaly Detection
Anomaly detection identifies unusual patterns that may signal errors, fraud, or attacks.
A
Answer Relevancy Metric
Answer Relevancy scores how well a generated answer addresses the input prompt.
Answer Relevancy RAG Metric
In RAG systems, this metric scores how well answers align with the asked question.
APPS Coding Benchmark
APPS evaluates how well LLMs solve programming problems with correct, efficient code.
A
ARC Reasoning Benchmark
ARC tests whether language models reason through grade-school science questions.
A
Argument Correctness Metric
This metric scores logical validity and structure of arguments produced by AI agents.
Artificial Neural Network
ANNs learn from data via interconnected units that approximate complex input–output relationships.
A
ASCII Smuggling Injection Attack
ASCII smuggling injects hidden instructions via invisible characters that models still parse.
Attention in Machine Learning
Attention lets models focus on the most relevant parts of an input when making predictions.
A
Attribute
Attributes are the data fields or features that models use to learn patterns and make predictions.
A
Auto-Encoders
Autoencoders learn compact representations by encoding data and decoding it back to the original.
A
Automated Machine Learning
AutoML streamlines model development by automating search over pipelines and hyperparameters.
A
AutoML
AutoML reduces manual ML work by automating model search, tuning, and basic evaluation.
A
Autonomous Agents
Autonomous agents perceive context, set goals, and act with little ongoing human control.
A
Autoregressive Model
Autoregressive models forecast future points from lagged past values in a sequence.
A
Average Precision
Average Precision averages precision across recall levels, often at multiple IoU thresholds for detection.
A
AWS Bedrock
AWS Bedrock provides API access to multiple foundation models with enterprise security features.
A
AWS Sagemaker
SageMaker provides IDE, training, and deployment tools for cloud machine learning on AWS.
A
Backpropagation
Backpropagation calculates gradients through a network so weights can be updated during learning.
B
Backpropagation Algorithm
Backpropagation adjusts weights in reverse through the network to minimize prediction error.
B
Bagging in Machine Learning
Bagging builds an ensemble from bootstrap samples to stabilize predictions and cut variance.
B
Baseline Distribution
Baseline distributions provide the minimum performance bar for evaluating advanced models.
B
Baseline Models
Baseline models are simple starting points used to judge whether complex models add real value.
B
Batch Normalization
Batch normalization stabilizes training by normalizing interlayer activations within mini-batches.
B
Batch Standardization
Batch standardization normalizes outputs between layers to mitigate internal covariate shift.
B
Bayes' theorem
Bayes' theorem computes conditional probability from priors and likelihoods.
B
Berkeley Function-Calling Leaderboard Domain-Specific Benchmark
Domain benchmark for multi-language LLM function calling, parallel calls, and relevance detection.
B
BERT
Bidirectional transformer encoder for contextual NLP representations and fine-tuning.
B
Best-of-N Prompt Injection Attack
Adversarial method that samples many prompt variants and keeps any that bypass injection defenses.
Bias Metric
Metric for political, gender, and social bias in model outputs used in decisions.
Bias Variance Tradeoff
ML tradeoff between underfitting bias and overfitting variance.
B
BigBench Reasoning Benchmark
Collaborative benchmark covering logic, math, and language comprehension for LLMs.
B
Binary Classification
Supervised task assigning each input to one of two classes.
B
Binary Cross Entropy
Log-loss metric for binary probability predictions versus true labels.
B
Binomial Distribution
Probability distribution of successes in n identical independent Bernoulli trials.
B
Black Box Model
Models whose internal decision mechanics are opaque to users and auditors.
B
BLEU
Automatic metric comparing generated text to references via n-gram precision.
B
Blue Teaming Capabilities (OWASP Taxonomy)
Defense capabilities: guardrails, runtime firewalls, AI-SPM, and detection that blocks or catches attacks.
B
Broken Function Level Authorization Excessive Agency Attack
Probe that tests whether agents execute functions beyond their authorized privilege level.
Broken Object Level Authorization Excessive Agency Attack
Probe that checks whether agents fetch unauthorized objects via prompts or tool calls.
Calibration Curve
Plot comparing predicted class probabilities to observed frequencies.
C
Canonical Schema
Standardized data model enabling consistent exchange across multiple systems.
C
Catastrophic Forgetting
Neural-net phenomenon where new learning overwrites previously acquired capabilities.
C
Catboost
Gradient boosting toolkit specializing in categorical feature handling.
C
Categorical Variables
Features that take values from a finite set of named categories.
C
Causal Language Modeling (CLM)
Autoregressive training that predicts the next token from left context only.
C
CBRN Harmful Content Attack
Adversarial probes seeking CBRN weapons-related assistance from AI models.
Chain-of-Thought
Intermediate step-by-step reasoning used by LLMs; also a security attack surface.
C
Chain-of-Thought Evaluation Metric
Metric that scores model outputs using systematic step-by-step reasoning criteria.
Chain-of-Thought Prompting
Prompting method that elicits intermediate reasoning steps for complex tasks.
C
Chatbot Arena Conversation Benchmark
Human preference benchmark ranking chat models with pairwise votes and Elo scores.
C
ChatGLM
Family of bilingual conversational LLMs focused on Chinese–English dialogue.
C
CI CD for Machine Learning
Applying continuous integration and delivery to ML training and deployment.
C
Citation Framing Injection Attack
Prompt injection that frames malicious asks as academic citations or scholarly references.
Classification Threshold
Cutoff on predicted probability used to assign a hard class label.
C
Class Imbalance
Training data where one class far outnumbers others, skewing naive metrics.
C
Clustering Algorithms
Unsupervised methods that group data points by similarity without labels.
C
CodeContests Coding Benchmark
Competitive-programming benchmark for assessing algorithmic code generation by LLMs.
C
Code Execution Metric
Metric that evaluates generated code by executing it against tests or expected outputs.
Code Interpreter
LLM tool that executes model-generated code, typically inside a sandbox.
C
CodeXGLUE Coding Benchmark
Microsoft multi-task benchmark suite for code understanding and generation models.
C
CommonsenseQA Reasoning Benchmark
Multiple-choice QA benchmark requiring everyday commonsense reasoning.
C
Competitors Brand Damage Attack
Probe that checks if AI can be pushed to endorse competitors or damage brand reputation.
Complex Event Processing
Real-time analysis of event streams to detect complex patterns and situations.
C
Computer Vision
AI field for interpreting visual data such as images and video.
C
Confusion Matrix in Machine Learning
Table summarizing classifier correct and incorrect predictions by class.
C
Context Compliance Harmful Content Attack
Multi-turn probe using fabricated conversation history to elicit harmful content.
Contextual Precision Metric
RAG metric for whether relevant context ranks above irrelevant chunks.
Contextual Recall Metric
RAG metric for whether necessary supporting information was retrieved.
Contextual Relevancy Metric
Metric scoring how relevant retrieved or supplied context is to the query.
Contextual Relevancy RAG Metric
RAG metric for on-topic retrieved documents before answer generation.
Context Window
Maximum token span an LLM can attend to in a single request.
C
Continuous Integration Model
CI practice applied to model training, testing, and validation pipelines.
C
Continuous Validation
Repeated post-deploy testing of models for quality, safety, and drift.
C
Contrastive Learning
Self-supervised method learning by comparing similar vs dissimilar sample pairs.
C
Conversational Agent
AI systems that hold multi-turn natural-language conversations with users.
C
Convex Optimization
Mathematical optimization where convexity yields tractable global solutions.
C
Convolutional Neural Networks
Neural nets with convolutional layers for spatial pattern recognition.
C
Copyright Violations Harmful Content Attack
Probe for unauthorized reproduction or distribution of copyrighted materials by AI.
Corrective RAG
RAG pattern that grades context and triggers corrective retrieval when needed.
C
Correctness Metric
Metric comparing model outputs to expected ground-truth answers.
Crescendo Harmful Content Attack
Incremental multi-turn strategy that steers models toward harmful content gradually.
Cross-Lingual Language Models
AI models that understand and generate text across multiple languages via shared representations.
C
Cross Session Leak Data Privacy Attack
Privacy probe testing whether sensitive data leaks across separate AI sessions.
Cross Validation Modeling
A model assessment method that tests on multiple data folds for more reliable generalization estimates.
C
Custom LLM Judge Metric
An LLM-as-judge evaluation method with custom criteria and prompts for subjective quality scoring.
C
CyberSecEval Harmful Content Attack
A red-team probe using CyberSecEval samples to check if models produce harmful cybersecurity content.
DAN Prompt Injection Attack
DAN-style prompt injection uses role-play to push models past safety protocols as an unrestricted assistant.
Data Annotation in AI
Labeling data samples so supervised machine learning models can learn from them.
D
Data Augmentation
Techniques that synthetically expand training sets with modified copies of existing samples.
D
Data Binning
A preprocessing method that maps continuous numbers into discrete bin intervals.
D
Data-Centric AI
An approach that improves AI by systematically raising data quality and coverage.
D
Data Cleaning
The process of fixing or removing inaccurate, duplicate, or malformed data before analysis.
D
Data Decomposition
Separating a time series into trend, seasonal, and residual components.
D
Data Flywheel
A loop where usage generates data that improves models and drives further usage.
D
Data Granularity
How fine-grained or aggregated the records in a dataset are.
D
Data Logging
Automated collection and storage of time-stamped measurements or events.
D
Data Mart
A subject-specific subset of a data warehouse for a department or business function.
D
Data Purification
Cleaning inaccurate, corrupt, duplicate, or missing data from datasets.
D
Data Science Platform
Integrated software for collaborative data prep, modeling, and insight sharing.
D
Data Science Techniques
Analytical and modeling methods used to extract actionable insights from data.
D
Data Science Tools
Software libraries and platforms used to collect, analyze, and model data.
D
Datasets And Machine Learning
Structured collections of examples that train, validate, and test ML models.
D
Data Vault
A data warehouse modeling approach built around hubs, links, and satellites.
D
Data Versioning
Tracking and labeling dataset states so ML experiments remain reproducible.
D
Data Visualizations
Graphical representation of data to reveal trends, anomalies, and patterns.
D
Decision Boundary
The region in feature space where a model's predicted class switches.
D
Decision Intelligence
A discipline that blends analytics, AI, and decision processes for better outcomes.
D
Decision Tree
A model that routes examples through feature-based splits to a leaf prediction.
D
Decision Tree In Machine Learning
Supervised models that learn hierarchical splits for classification or regression.
D
Decomposed Evaluation Metric
An evaluation approach that scores separate quality dimensions instead of one opaque overall grade.
D
Deep Belief Networks
Generative deep models built by stacking restricted Boltzmann machines.
D
Deep Learning
Neural network methods with many layers that learn hierarchical representations.
D
Deep learning Algorithms
Neural architectures and training methods used to build deep learning systems.
D
Deep Q-Network
A reinforcement learning method that uses deep nets to approximate action values.
D
Deep Reinforcement Learning
Using deep neural networks as function approximators inside reinforcement learning.
D
Deepset Injection Attack
A prompt-injection evaluation that probes whether models follow attacker instructions over system intent.
Deep SHAP
A SHAP-based method for explaining predictions of deep neural networks.
D
Degradation Model
Statistical or ML models of how performance decays under time or stress.
D
Denial of Service Attack
DoS against AI systems via long outputs, loops, or resource-heavy prompts that exhaust capacity or spend.
DenseNET
A convolutional architecture with dense skip connections between layers.
D
Density Based Clustering
Clustering that finds dense regions of arbitrary shape and marks outliers as noise.
D
Diffusion Models
Generative models that create samples by reversing a stepwise noising process.
D
Dimensionality Reduction
Techniques that map high-dimensional data into a lower-dimensional representation.
D
Direct Preference Optimization
An alignment method that trains LLMs from human preference pairs without a separate reward model.
D
Divergent Repetitions Training Data Extraction Attack
An extraction probe that uses divergent repetition patterns to surface memorized or sensitive training content.
DoNotAnswer Harmful Content Attack
A harmful-content probe based on DoNotAnswer-style questions that safe AI systems should refuse.
Dplyr
An R package for expressive, pipe-friendly tabular data manipulation.
D
Drift Monitoring
Tracking distribution or behavior shifts in data and models against a known baseline.
D
DROP Reasoning Benchmark
A QA benchmark testing discrete reasoning skills over paragraph-length contexts.
D
Early Stopping
A regularization tactic that stops training when held-out metrics plateau or worsen.
E
Embedding Projector
Interactive tools for projecting and exploring embedding spaces visually.
E
Embedding Similarity Metric
A metric that scores how close two texts are in embedding space for semantic evaluation.
Embeddings in Machine Learning
Dense vector representations that encode semantic or structural similarity for ML tasks.
E
Embodied Agents
AI agents that perceive and act through a physical or simulated body, not text alone.
E
Emotion Detection Metric
A metric that classifies or scores emotional tone in model outputs for UX and safety reviews.
Encoding Prompt Injection Attack
Prompt injection that obfuscates harmful instructions via encodings like Base64 or NATO phonetic text.
End-to-End Evaluation
Testing the complete AI workflow from user input through tools and final response quality.
E
Ensemble Learning
Methods that aggregate multiple models to produce a stronger predictor.
E
Enterprise Generative AI
Organization-scale generative AI with governance, security controls, and continuous evaluation.
E
Epoch in machine Learning
One full pass of the training set through the learning algorithm.
E
Ethical AI
Practices that align AI systems with fairness, transparency, privacy, and human rights.
E
ETL Pipeline
Extract–transform–load workflow moving data from sources into analytics systems.
E
Evolutionary Algorithms
Population-based optimizers that evolve candidate solutions via selection and variation.
E
Exact Match Metric
A precise evaluation method that scores answers via direct string equality to gold labels.
Explainable AI (XAI)
AI designed to provide understandable insights into how and why decisions are made.
E
Exploratory Data Analysis (EDA)
Foundational analysis that reveals dataset structure, often with visualizations.
E
F1-Score Metric
Harmonic mean of precision and recall for evaluating classification and extraction quality.
Facial Recognition
Biometric tech that matches faces in images or video to known identities.
F
Failure Analysis Machine Learning
Systematic diagnosis of why machine learning models make incorrect predictions.
F
Faithfulness Metric
Scores how well an answer stays faithful to retrieved or provided source context.
False Positive Rate
Proportion of actual negatives incorrectly classified as positive.
F
Feature Engineering
Creating and transforming input variables to improve machine learning models.
F
Feature Selection
Choosing the most informative subset of features for a model.
F
Feature Vector
An ordered set of numeric features representing a single data example.
F
Federated Learning
Collaborative training that keeps raw data local and aggregates model updates.
F
Feedback Loop
When model predictions alter future data and can amplify bias or error.
F
Few-Shot Evaluation Metric
Assesses model performance when learning from a small number of in-context examples.
Few-Shot Learning
ML paradigm where models adapt to new tasks from only a few labeled examples.
F
Few-Shot Prompting
Prompting technique that includes a few demonstrations to guide LLM behavior.
F
FinBen Domain-Specific Benchmark
A domain benchmark for assessing LLM performance on financial tasks.
F
Folium
Python wrapper for creating interactive Leaflet maps from geospatial data.
F
F-score
A performance metric combining precision and recall into a single score (often F1).
F
GAIA Agent Benchmark
A benchmark of real-world questions for evaluating general AI assistants and agents.
G
Gaussian Distribution
The normal bell-curve distribution parameterized by mean and variance.
G
Gaussian Mixture Model
Probabilistic model representing data as a mixture of Gaussian components.
G
GCG Injection Harmful Content Attack
Greedy Coordinate Gradient (GCG) attacks optimize adversarial suffixes that jailbreak LLM safety filters.
Generalist Language Model
An LLM designed for broad tasks across domains rather than one specialized use case.
G
Generalized Linear Models
A statistical family linking linear predictors to diverse response distributions.
G
Generative Adversarial Networks
Two-network architecture where a generator learns by fooling a discriminator.
G
Generative Agents
LLM-based agents that simulate believable behavior via memory, planning, and reflection.
G
G-Eval
An LLM-as-judge method that scores outputs via structured evaluation rubrics.
G
GOAT Attack Harmful Content Attack
GOAT automates multi-turn adversarial prompting to uncover jailbreaks and harmful-content failures in LLMs.
Gradient Boosting
Sequential ensemble method that adds weak learners to correct prior residuals.
G
Gradient Descent in Machine Learning
Iterative optimization that follows gradients to minimize training loss.
G
Grandma Framing Injection Attack
Emotional framing that wraps harmful asks as innocent family scenarios to jailbreak LLMs.
Grid Search
Brute-force search over a specified hyperparameter grid.
G
Groundedness Metric
Evaluates whether answer claims are supported by provided or retrieved context.
Ground Truth
Authoritative reference data against which model predictions are scored.
G
Grouped Query Attention
Transformer attention variant that groups queries to reduce KV cache size.
G
GSM8K Math Benchmark
Benchmark of grade-school math word problems for multi-step reasoning.
G
Hallucination Index
A composite score for how often an LLM fabricates facts without support.
H
Hallucination Metric
An evaluation measure for unsupported or fabricated claims in LLM outputs.
Handling Outliers
Methods to detect and treat extreme values that distort analysis or training.
H
HarmBench Harmful Content Attack
A benchmark suite for measuring LLM resistance to harmful content generation.
Hash tables
Fast key–value maps implemented with hashing and collision resolution.
H
Hash Tables
Data structure mapping keys to values via hashing for fast average-case lookup.
H
HellaSwag Reasoning Benchmark
A multiple-choice benchmark for commonsense natural language inference.
H
Hellinger Distance
A metric for comparing probability distributions, related to Bhattacharyya coefficient.
H
Helpfulness Metric
Scores how useful and goal-aligned an LLM response is for the user request.
Hijacking Excessive Agency Attack
An attack that manipulates an AI agent into acting beyond its intended permissions or tool scope.
Holdout Data
A train/test split portion reserved solely for unbiased model evaluation.
H
HumanEval
A benchmark of programming problems scoring LLM-generated code via unit tests.
H
HumanEval Coding Benchmark
Functional-correctness benchmark for LLM code synthesis using unit tests.
H
Human-in-the-Loop Machine Learning
ML workflows that integrate human judgment for labeling, feedback, or oversight.
H
Hyperparameter Optimization
Systematic search for configuration values that maximize validation performance.
H
Hyperplane
A flat (n−1)-dimensional subspace used as a linear decision boundary in ML.
H
Illegal Activities Harmful Content Attack
Red-team probe for prompts that push AI toward illegal or harmful criminal content.
Image Data Augmentation
Artificially expanding image datasets with transforms to improve vision model generalization.
I
Image Data Collection
Gathering and organizing images for training computer vision models.
I
Imbalanced Data
Datasets where class frequencies are highly uneven, biasing naive classifiers.
I
Impersonation Brand Damage Attack
Probe for AI impersonation of people or brands that could damage reputation.
In-Context Learning
Adapting an LLM to a task via examples in the prompt rather than fine-tuning.
I
Independent and Identically Distributed Data IID
Data where samples are independent and share the same distribution.
I
Information Retrieval
Systems that find and rank documents relevant to a user query.
I
Instruction Tuning
Fine-tuning LLMs on instruction–response pairs to improve instruction following.
I
Intelligent Document Processing IDP
OCR + NLP + ML pipelines that extract structured data from documents.
I
Intent Classification Metric
A metric that checks whether text is mapped to the correct intent label.
Intersection over Union (IoU)
Overlap ratio between predicted and ground-truth bounding boxes.
I
JSON Validation Metric
Automated checks that generated JSON is valid and includes required keys.
Juries of Models Metric
Evaluating outputs with multiple LLM judges for more stable scores.
J
Keras
A user-friendly high-level API for defining and training neural networks.
K
K-Nearest Neighbor KNN
A lazy supervised algorithm that classifies via the K closest labeled examples.
K
KNN Models
Non-parametric models that predict from the K nearest training examples.
K
Kolmogorov-Smirnov Test
A nonparametric test comparing sample distributions via maximum CDF distance.
K
KYC Process
Identity verification processes that deter fraud and meet financial regulations.
K
LangChain
Open-source framework for building LLM chains, agents, and tools.
L
Language Classification Metric
A metric that detects which natural language a text sample is written in.
Large Action Models
Models that map natural-language goals into executable actions in digital environments.
L
Learning Rate in Machine Learning
A hyperparameter controlling how much weights update on each training step.
L
Learning-to-Rank
Supervised ML that optimizes the order of search or recommendation results.
L
LegalBench Domain-Specific Benchmark
A 162-task benchmark covering six types of legal reasoning for AI evaluation.
L
Liability Engagement Legal Risk Attack
Probe for AI agents making commitments that create legal or financial liability.
LightGBM
A leaf-wise gradient boosting library optimized for speed on tabular data.
L
Likert Framing Injection Attack
Jailbreak that frames harmful queries as Likert-scale research surveys.
Linear Regression
A statistical method modeling a dependent variable as a linear combination of predictors.
L
Link Injection Data Privacy Attack
Probe for AI generating phishing, malware, or other malicious links.
Llama
Meta’s open LLM family used for research and many commercial applications.
L
LlamaIndex
Framework connecting datasets to LLMs for context-aware RAG apps.
L
LLM Agents
Language-model systems that plan, call tools, and act toward user goals.
L
LLM Alignment
Steering LLMs so behavior matches human values, intent, and safety policies.
L
LLM APIs
HTTP interfaces that let applications call large language models for text tasks.
L
LLM-as-a-Service
Cloud-hosted access to large language models via APIs, without operating your own training or inference clusters.
L
LLM Benchmarks
Standardized task suites for comparing language model performance.
L
LLM Chatbot Evaluation
Systematic testing of chatbot quality, safety, and user experience.
L
LLM Cost
Token, infrastructure, and routing costs of operating large language models.
L
LLM Debugger
Tools and practices for diagnosing LLM prompt and output failures.
L
LLM Deployment
Integrating LLMs into production with performance, safety, and ops controls.
L
LLM Distillation
Training a compact model to mimic a larger LLM’s outputs and skills.
L
LLM Embeddings
Dense vector representations of text used for search and retrieval.
L
LLM Evaluation
Systematic measurement of LLM accuracy, safety, and usefulness.
L
LLM Evaluation Framework
Structured protocol for assessing LLM performance, reliability, and ethics.
L
LLM Fine-Tuning
Further training a pre-trained LLM on specialized data for niche tasks.
L
LLM Gateway
A proxy layer for routing, auth, and policy across LLM providers.
L
LLM Guardrails
Runtime rules and filters that keep LLM apps safe, compliant, and secure.
L
LLM Hallucinations
When LLMs produce fluent but false or unsupported statements.
L
LLM Inference
Using a trained LLM to generate outputs on new, unseen inputs.
L
LLM Interpretability
Methods that explain how LLMs turn inputs into outputs.
L
LLM Jailbreaking
Bypassing LLM safety restrictions through adversarial prompts and multi-turn attacks.
L
LLM Knowledge Base
Curated content used to ground LLM answers in trusted sources.
L
LLM Knowledge Graph
Entity-relation graphs used to ground LLM answers with structured facts.
L
LLM Leaderboards
Public rankings of language models on shared benchmark suites.
L
LLM Observability
Monitoring prompts, traces, and quality signals for LLM applications.
L
LLM Ontology
Concept hierarchies that structure domain knowledge for LLM reasoning.
L
LLMOps
Practices and tooling for deploying, monitoring, evaluating, and governing large language models in production systems.
L
LLM Orchestration
Coordinating models, tools, and memory into multi-step AI workflows.
L
LLM Output Parsing
Extracting structured, machine-usable data from free-form LLM responses for safe integration with tools and APIs.
L
LLM Overreliance
Excessive dependence on LLM outputs for decisions that still need human judgment, verification, and domain expertise.
L
LLM Parameters
Model weights and decoding settings that control how large language models interpret prompts and generate outputs.
L
LLM Playground
An interactive environment for experimenting with LLM prompts, models, and parameters before production integration.
L
LLM Product Development
Building products that embed large language models into user-facing workflows with clear value, evaluation, and iteration loops.
L
LLM Quantization
Reducing the numeric precision of model weights (and sometimes activations) to shrink size and speed inference with minimal quality loss.
L
LLM Red Teaming
Adversarial testing of large language models to find jailbreaks, data leaks, bias, and harmful outputs before deployment.
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.
L
LLM Stack Layers
The layered toolchain for LLM apps: data, model, orchestration, serving, evaluation, and interface components.
L
LLM Summarization
Using large language models to condense long documents into shorter text while preserving key information.
L
LLM Testing
Systematic evaluation of large language models for accuracy, robustness, safety, and reliability before production use.
L
LLM Toxicity
Hate, harassment, or abusive content produced by large language models, often learned from unfiltered training data.
L
LLM Tracing
Recording prompts, responses, tool calls, latency, and costs across LLM application runs for debugging and observability.
L
Locally Interpretable Model-Agnostic Explanations LIME
A model-agnostic method that explains individual predictions with a simple local surrogate model around the instance.
L
Logistic Regression
A classification method that models the probability of categorical outcomes using a logistic (sigmoid) function.
L
Long Short Term Memory LSTM
A recurrent neural network architecture that retains long-range sequence information via gated memory cells.
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.
L
Machine Learning
A field of AI where algorithms learn patterns from data to make predictions or decisions without being fully hard-coded.
M
Machine Learning Algorithm
A computational method that learns patterns from data to predict or decide on unseen examples.
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.
M
Machine Learning Bias
Systematic skew in ML predictions that unfairly favors or disadvantages groups, often from biased training data or objectives.
M
Machine Learning Checkpointing
Periodically saving model and optimizer state during training to enable recovery, comparison, and rollback.
M
Machine Learning Deployment
Integrating trained ML models into production applications and APIs so they serve live predictions.
M
Machine Learning Inference
Executing a trained model on new data to produce predictions or generations outside the training loop.
M
Machine Learning in Software Testing
Using ML to generate, prioritize, or analyze software tests for better coverage and faster defect detection.
M
Machine Learning Lifecycle
End-to-end stages of an ML project: problem framing, data, training, validation, deployment, and monitoring.
M
Machine Learning Model Accuracy
The proportion of correct predictions; often paired with precision, recall, and slice-aware metrics for decisions.
M
Machine Learning Model Evaluation
Assessing ML model effectiveness with held-out data and metrics like accuracy, AUC, F1-score, and calibration.
M
Machine Learning Orchestration
Automating and coordinating ML pipeline stages—from data prep to deployment and monitoring—at scale.
M
Machine Learning Pipeline
A structured workflow from data preparation through training, evaluation, deployment, and monitoring of ML models.
M
Machine Learning Workflows
Ordered steps for delivering ML solutions: data collection, preparation, training, evaluation, and iteration.
M
MAP (Mean Average Precision) Metric
A ranking metric that averages precision across queries, widely used for search and retrieval evaluation.
M
Math Framing Injection Attack
A jailbreak pattern that hides harmful intents inside mathematical or academic framing to bypass AI safety filters.
MATH Math Benchmark
A benchmark of competition-style math problems used to evaluate LLM reasoning from elementary through high-school topics.
M
MBPP Coding Benchmark
A Python coding benchmark with hundreds of problems assessing correctness and core programming skill in generated solutions.
M
Mean Absolute Error
Average absolute gap between predictions and true values; a common regression accuracy metric.
M
Mean Absolute Percentage Error
Average absolute percentage difference between predicted and actual values; common in forecasting evaluation.
M
Mean Squared Error (MSE)
Average of squared prediction errors; a common regression loss and accuracy metric.
M
Memory-Augmented Neural Networks
Neural networks extended with external memory for longer-term storage and retrieval during complex tasks.
M
Meta Learning
A paradigm where algorithms learn how to learn, enabling fast adaptation to new tasks from limited examples.
M
METEOR Score
A translation evaluation metric that aligns outputs to references using stems and synonyms, balancing precision and recall.
M
Micro-Models
Compact models focused on narrow tasks or data slices within a larger modeling system.
M
Misinformation & Disinformation Harmful Content Attack
An adversarial probe that tries to make AI agents generate or amplify false, misleading, or harmful information.
Missing Values in Time Series
Absent observations in sequential datasets that require removal, imputation, or specialized models to avoid biased analysis.
M
Mixture of Experts
An architecture that routes each input to a small subset of expert subnetworks for scalable model capacity.
M
ML Architecture
The blueprint for how data, models, serving, and monitoring components connect in an ML system.
M
ML-Based Scoring Metric
Cost-effective text evaluation with pre-trained scorers instead of LLM-as-judge.
M
ML Diagnostics
Systematic identification and remediation of performance, bias, and robustness issues in ML models.
M
ML Infrastructure
Compute, storage, and tooling foundations for building, training, and deploying machine learning models.
M
ML Interpretability
Understanding and describing how ML models make predictions, beyond raw accuracy scores.
M
ML Model Cards
Standardized documentation of an ML model’s use, data, metrics, biases, and limitations.
M
ML Model Management
Systems and practices for versioning, approving, deploying, and monitoring machine learning models.
M
ML Model Validation
Confirming that models meet quantitative and qualitative requirements on intended data before deployment.
M
MLOps
Practices that connect ML development and operations for scalable, governed deployments.
M
MLOps for Generative AI
CI/CD, monitoring, and governance adapted for generative AI systems and agents.
M
MLOps Framework
A structured methodology for automating and governing the ML lifecycle end to end.
M
MLOps Monitoring
Tracking live ML performance, drift, and feedback so production models stay trustworthy.
M
ML Performance Tracing
Monitoring model predictions and operational metrics across the lifecycle to find bottlenecks and quality drift.
M
ML Scalability
Ability of an ML system to process more data or traffic without unacceptable loss of performance or precision.
M
ML Stack
The collection of software tools and frameworks used to develop, deploy, and analyze machine learning solutions.
M
MMLU benchmark
Massive Multitask Language Understanding: a broad academic benchmark for LLM capability.
M
MMLU Reasoning Benchmark
Multiple-choice academic questions used to probe an LLM knowledge and reasoning breadth.
M
Model-Based Machine Learning (MBML)
Tailoring probabilistic models to problems via factor graphs and Bayesian inference.
M
Model Behavior
How a model predicts and behaves under real inputs, shaped by data, structure, and training.
M
Model Calibration
Adjusting predicted probabilities so they match observed outcome frequencies.
M
Model Distillation
Compressing a large teacher model into a smaller student that retains most of its behavior.
M
Model Drift
Performance loss when live data or conditions diverge from what the model was trained on.
M
Model-Driven Architecture
Software development centered on platform-independent models and transformations.
M
Model Explainability
Methods that make ML decisions understandable to stakeholders and auditors.
M
Model Fairness
Ensuring ML predictions avoid unjustified disparate impact across groups.
M
Model Merging
Combining multiple models weights or adapters into one stronger or more specialized model.
M
Model Monitoring
Real-time observation of ML performance to detect failures and meet business requirements.
M
Model Observability
Inspecting ML inputs, outputs, and internals to debug and optimize production systems.
M
Model Parameters
Learned internal coefficients that determine how a model transforms inputs into predictions.
M
Model Registry
Centralized versioning and lifecycle management for production-ready ML models.
M
Model Retraining
Updating a trained model with new data to improve performance or adapt to change.
M
Model Robustness
A model's ability to stay accurate under noise, shifts, and adversarial conditions.
M
Model Selection
Choosing the best model among candidates using validation and resampling techniques.
M
Model Tuning
Adjusting hyperparameters to optimize a model's learning process and accuracy.
M
MRR (Mean Reciprocal Rank) Metric
Ranking metric focused on how soon the first correct answer appears.
M
MT-Bench
A multi-turn benchmark for conversational LLM quality and coherence.
M
MT-Bench Conversation Benchmark
Benchmark for how well LLMs stay coherent across conversation turns.
M
MTEB
Massive Text Embedding Benchmark for comparing embedding model quality.
M
Multi-Class Classification
Predicting one class among three or more possible labels for each instance.
M
Multilayer Perceptron (MLP)
Feedforward neural net with one or more hidden layers for classification and regression.
M
Multilingual LLM
Large language models trained to understand and generate text in many languages.
M
MultiMedQA Domain-Specific Benchmark
Medical QA benchmark combining datasets to score factuality, harm, and bias.
M
Naive Bayes Model
Fast probabilistic classifier using Bayes theorem with independent-feature assumptions.
N
Natural Language Search
Search that understands user intent from free-form language, not just keyword match.
N
Natural Language Understanding
AI methods that let machines interpret meaning and intent in human language.
N
NDCG (Normalized Discounted Cumulative Gain) Metric
Ranking metric that rewards relevant items higher up the list, normalized to an ideal order.
N
Neural Networks
Layered models of artificial neurons used for learning patterns from data.
N
Neural Network Tuning
Optimizing neural net hyperparameters like depth, learning rate, and batch size.
N
No-code Low-code ML
Platforms that build ML apps and workflows with visual tools instead of heavy coding.
N
Noise in Machine Learning
Unwanted data variation that can mislead training and degrade generalization.
N
Noisy Image
Images with unwanted brightness or color variation that may need denoising.
N
Normalization in Machine Learning
Rescaling features to a common range (e.g., min-max) for stable model training.
N
Normalized Discounted Cumulative Gain
Ranking effectiveness metric balancing relevance and position, normalized to ideal order.
N
Nvidia NIM
NVIDIA's inference microservices for deploying optimized AI models at scale.
N
Observation ML
A single data instance (feature vector) analyzed or predicted by a learning algorithm.
O
One-Hot Encoding
Representing categories as binary columns so models can process categorical features.
O
Online Machine Learning
Incremental learning from streaming data so models adapt without full batch retrains.
O
Open-Source Machine Learning Monitoring
Community tools for supervising live ML models—drift, performance, and alerts.
O
Out-of-distribution
Inputs that differ substantially from training data, often causing unreliable predictions.
O
Overfitting in Machine Learning
Learning training noise so closely that performance on new data drops.
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.
O
OWASP ASI01 Agent Goal Hijack
When email, PDFs, or tool output rewrite an agent objective so the workflow chases the wrong goal.
A
OWASP ASI02 Tool Misuse and Exploitation
When agents misuse real tools and permissions after injection or misalignment.
T
OWASP ASI03 Identity and Privilege Abuse
Delegation chains and inherited credentials that escalate access without re-checking the user.
I
OWASP ASI04 Agentic Supply Chain Vulnerabilities
Runtime-loaded MCP servers, prompt packs, and agent personas you did not fully trust.
A
OWASP ASI05 Unexpected Code Execution
Coding agents and vibe-coding tools running shell commands nobody reviewed.
U
OWASP ASI06 Memory and Context Poisoning
Bad facts that persist across sessions and compound every future task.
M
OWASP ASI07 Insecure Inter-Agent Communication
ASI07: weak auth, spoofed peers, and unencrypted A2A/MCP traffic between agents.
I
OWASP ASI08 Cascading Failures
ASI08: one poisoned agent output cascading through a multi-agent workflow before humans can stop it.
C
OWASP ASI09 Human-Agent Trust Exploitation
ASI09: agents using urgency or authority cues to push humans into unsafe approvals.
H
OWASP ASI10 Rogue Agents
ASI10: misaligned or self-persisting agents that keep acting after they should be revoked.
R
OWASP LLM01 (2023) Prompt Injection
The 2023/24 OWASP #1 LLM risk: crafted inputs that override developer instructions.
P
OWASP LLM01 (2025) Prompt Injection
OWASP #1 LLM risk: direct, indirect, and automated prompt injection that overrides instructions.
P
OWASP LLM02 (2023) Insecure Output Handling
2023 OWASP risk: apps trusting LLM output into browsers, shells, or APIs without sanitization.
I
OWASP LLM02 (2025) Sensitive Information Disclosure
When LLMs leak PII, credentials, or internal context through chat, tools, or cross-session memory.
S
OWASP LLM03 (2023) Training Data Poisoning
2023 OWASP entry for poisoned training or fine-tuning data that embeds backdoors or bias.
T
OWASP LLM03 (2025) Supply Chain
LLM supply-chain risk from models, plugins, datasets, and MCP servers you did not build.
S
OWASP LLM04 (2023) Model Denial of Service
2023 OWASP risk: resource exhaustion that degrades LLM availability for legitimate users.
M
OWASP LLM04 (2025) Data and Model Poisoning
Poisoned training, fine-tuning, RAG, or memory that steers trusted LLM outputs wrong.
OWASP LLM05 (2023) Supply Chain Vulnerabilities
2023 OWASP risk from untrusted models, datasets, plugins, and weights in the LLM stack.
S
OWASP LLM05 (2025) Improper Output Handling
When trusted LLM text becomes XSS, SQL injection, or shell execution in downstream systems.
I
OWASP LLM06 (2023) Sensitive Information Disclosure
2023 OWASP entry for LLMs revealing confidential data in responses or context.
S
OWASP LLM06 (2025) Excessive Agency
When an LLM can call APIs, write data, or send messages without enough human checks.
E
OWASP LLM07 (2023) Insecure Plugin Design
2023 OWASP risk from poorly designed LLM plugins that escalate privilege or trust input.
I
OWASP LLM07 (2025) System Prompt Leakage
When crafted prompts extract system instructions, guardrails, or tool schemas.
S
OWASP LLM08 (2023) Excessive Agency
2023 OWASP risk: LLM agents granted more autonomy and tools than safety controls justify.
E
OWASP LLM08 (2025) Vector and Embedding Weaknesses
RAG attack surface: poisoned chunks, weak vector DB access, and manipulated similarity.
V
OWASP LLM09 (2023) Overreliance
2023 OWASP risk: people and systems trusting fluent LLM answers without verification.
O
OWASP LLM09 (2025) Misinformation
Fluent, authoritative LLM falsehoods that drive bad decisions in finance, legal, or medical use.
M
OWASP LLM10 (2023) Model Theft
2023 OWASP risk of stealing model weights or approximating behavior through queries.
M
OWASP LLM10 (2025) Unbounded Consumption
Runaway tokens, tool loops, and inference spend that drain budget without a hard crash.
U
Pandas and NumPy
Pandas and NumPy are core Python libraries for tabular analysis and fast numerical array computation.
P
Parameter-Efficient Fine-Tuning
PEFT adapts LLMs by updating a small parameter set (e.g. LoRA) instead of full fine-tuning.
P
Parameter-Efficient Fine-Tuning (Prefix-Tuning)
Prefix-tuning learns continuous prefix vectors while freezing most base LLM weights.
P
Pascal
PASCAL VOC is a classic object detection and segmentation benchmark with standardized bounding boxes.
P
Pattern Matching
Pattern matching finds exact or regex-defined patterns in sequences of data or tokens.
P
Pattern Recognition
Pattern recognition identifies regularities in data so systems can classify or detect structures.
P
Permutation Importance
Permutation importance estimates feature influence by measuring performance drop after shuffling a feature.
P
Phare Safety Benchmark
Phare is a multilingual LLM safety benchmark for probing harmful content and refusal behavior.
P
PII Detection Metric
A PII detection metric scores whether model outputs expose personally identifiable information.
PII Leak Data Privacy Attack
A PII leak attack probes whether an LLM reveals personally identifiable information it should withhold.
Politeness Metric
A politeness metric scores whether model responses stay courteous and appropriately toned.
Politics Topic Harmful Content Attack
Safety probe checking whether an LLM refuses disallowed political harm content.
Pooling Layers in CNN
Pooling layers downsample CNN feature maps to reduce spatial size and keep salient activations.
P
Population Stability Index
PSI measures how much a feature distribution shifted between a reference and current population.
P
Positional Encoding
Positional encoding injects sequence-order information into transformer inputs that lack position.
P
Precision in Machine Learning
Precision is the fraction of predicted positives that are actually positive.
P
Precision Metric
A precision metric scores how many of the systems claimed positives are correct.
Precision-Recall Area Under the Curve (PR AUC)
PR AUC summarizes the precision-recall curve, especially useful on imbalanced classification problems.
P
Predictive Model Validation
Predictive model validation estimates generalization via holdout sets, cross-validation, or backtests.
P
Preprocessing
Preprocessing transforms raw inputs into features models can learn from.
P
Pre-trained Transformer
A pre-trained transformer is trained on large corpora then adapted downstream via fine-tuning or prompting.
P
Principal Component Analysis (PCA)
PCA projects data onto orthogonal components that capture maximal variance.
P
Probabilistic Classification
Probabilistic classification predicts class probabilities, not only hard labels.
P
Programmatic Validation Metric
A programmatic validation metric scores outputs with deterministic rules such as schemas or regex.
Prompt Alignment Metric
A prompt alignment metric scores how well a response follows user instructions and constraints.
Prompt Chaining
Prompt chaining sequences multiple LLM calls so each steps output feeds the next stage.
P
Prompt Engineering
Prompt engineering designs instructions and context that steer LLM behavior toward reliable outcomes.
P
Prompt Extraction Internal Information Attack
A prompt extraction attack tries to coerce an LLM into revealing its system prompt or hidden instructions.
Prompt Injection
Prompt injection occurs when untrusted text overrides developer instructions and steers model behavior.
P
Prompt Playground
A prompt playground is an interactive UI for trying prompts, models, and parameters during development.
P
Prototype Model
A prototype model is an early proof-of-concept trained to validate feasibility before production hardening.
P
Purple Teaming Capabilities (OWASP Taxonomy)
OWASP purple teaming: map red findings to blue controls and measure what actually works.
P
PyTorch
PyTorch is an open-source deep learning framework with dynamic graphs and pythonic tensor APIs.
P
RAG Architecture
System design that retrieves external context and feeds it into a generative model for grounded answers.
R
RAGAS
An evaluation framework for measuring RAG pipeline quality and reliability.
R
RAG as a Service
Cloud-managed RAG that packages retrieval, indexing, and generation for applications.
R
RAGET Complex Question Hallucination Attack
Uses complex questions derived from your knowledge base to test RAG hallucination resistance.
RAGET Simple Question Hallucination Attack
Generates simple factual questions from your knowledge base to evaluate RAG answer accuracy.
RAGET Situational Question Hallucination Attack
Tests RAG hallucinations using situational and user-specific question framing drawn from your docs.
RAG Evaluation
How to assess RAG systems across retrieval quality and generated answer faithfulness.
R
RAG Hallucinations
When retrieval-augmented models still produce false or unsupported claims that sound credible.
R
Random Forest
Ensemble of decision trees using bagging and random feature selection.
R
Random Initialization
Initializing network weights randomly near zero to break symmetry for gradient descent.
R
Readability Assessment Metric
Scores how easy AI-generated text is for a target audience to read.
Reasoning Engine
Component that performs structured inference or planning to produce conclusions or actions.
R
Recall in Machine Learning
Fraction of true positives identified among all actual positive cases.
R
Recall Metric
The share of true positives found among all actual positives.
Recall-Oriented Understudy for Gisting Evaluation (ROUGE)
N-gram recall metrics for comparing generated text to reference summaries.
R
Receiver Operating Characteristic (ROC) Curve
Plot of TPR vs FPR used to assess binary classifier discrimination.
R
Rectified Linear Unit (ReLU)
Activation that zeros negatives and passes positives unchanged.
R
Recurrent Neural Networks
Neural nets with temporal state for sequential data.
R
Red Teaming Capabilities (OWASP Taxonomy)
OWASP taxonomy red-teaming capabilities: adversarial probes, multi-turn attacks, and continuous offensive testing for AI systems.
R
Reference Distribution
Baseline distribution against which samples or model outputs are compared.
R
Regex Pattern Matching Metric
Deterministic checks that match or reject outputs using regular expressions.
Regression
Supervised learning that estimates continuous outcomes.
R
Regression Algorithms
Family of methods that map features to continuous targets.
R
Regularization Algorithms
Techniques that penalize complexity to reduce overfitting.
R
Regularization in Machine Learning
Constraining model complexity so predictors generalize to new data.
R
Reinforcement learning
Learning optimal actions via rewards and penalties over time.
R
Reinforcement Learning from AI Feedback (RLAIF)
Aligning models with preference signals produced by AI judges.
R
Relevance Metric
Scores how well results or answers match the user's information need.
Religion Topic Harmful Content Attack
Adversarial prompts that provoke biased or harmful statements about religion to test fairness.
Reproducible AI
Practices that make AI experiments and evaluations repeatable across runs and teams.
R
ResNet
CNN architecture with residual connections enabling deeper training.
R
Responsible AI
Principles and practices for building AI that is safe, fair, transparent, and accountable.
R
Retrieval-augmented Generation (RAG)
LLM pattern that retrieves external knowledge before generating a response.
R
Ridge Regression
Linear model with squared-coefficient penalty for stability.
R
RMSprop
Optimizer that scales steps by recent gradient magnitude.
R
Robotic Process Automation (RPA)
Automating structured business processes with software robots.
R
ROC-AUC Metric
Area under the ROC curve; measures ranking quality independent of a single threshold.
R
Root-Cause Analysis
Structured investigation to find underlying causes of failures.
R
Root Mean Square Error (RMSE)
Square root of mean squared error between predictions and targets.
R
Rotating Proxy
Proxy setup that cycles IP addresses across requests.
R
SafetyBench Safety Benchmark
A benchmark suite for measuring large language model safety performance.
S
Safety Metric
Evaluation score for how safely an LLM refuses or handles harmful requests.
Scikit-learn
Python toolkit for traditional ML algorithms and evaluation.
S
Script Injection Data Privacy Attack
Tests whether AI responses can include harmful scripts that enable XSS and data privacy breaches.
Segmentation in Machine Learning
Grouping data into meaningful segments for modeling or CV.
S
Selective Sampling
Choosing informative samples for labeling to reduce annotation cost.
S
Self-Consistency Evaluation Metric
Measures agreement across multiple sampled answers or chain-of-thought paths.
S
Semantic Router
Routes queries to tools or agents using semantic similarity or classifiers.
S
Semi-supervised Learning
Learning from scarce labels combined with unlabeled examples.
S
Sensitivity and Specificity in ML
TPR and TNR metrics describing binary detection performance.
S
Sentiment Analysis
NLP task that labels text as positive, negative, neutral, or by emotion.
S
Sentiment Analysis Metric
Measures accuracy or agreement of sentiment/polarity predictions.
Seq2Seq Model
Encoder-decoder architectures transforming one sequence into another.
S
Shadow Deployment
Serving a candidate model on production traffic for comparison without user impact.
S
Shapley Values
Game-theoretic attributions explaining each feature's contribution.
S
Shell Injection Excessive Agency Attack
Probes for shell injection when AI agents can execute system commands - a key excessive-agency risk.
Six-Month Moratorium
2023 proposal to pause training of frontier models beyond GPT-4-class systems.
S
Sliding Window Attention
Attention pattern restricting each token to a fixed-size local window for efficiency.
S
Softmax Function
Normalizes logits into a probability distribution summing to one.
S
SQL Injection Excessive Agency Attack
Tests whether AI agents with DB access can be prompted into unauthorized SQL.
SSRF Excessive Agency Attack
Tests AI agents for prompt-induced SSRF through browsing or HTTP tools.
Stereotypes and Discrimination Harmful Content Attack
Tests whether models resist prompts that push stereotyped or discriminatory outputs.
StrategyQA Reasoning Benchmark
Benchmark of yes/no questions needing implicit multi-step reasoning strategies.
S
Summarization Metric
Evaluates summary factual alignment and coverage of essential source content.
Supervised Machine Learning
ML paradigm that learns a mapping from features to labels using labeled training data.
S
Support Vector Machines (SVM)
Supervised models that separate classes with a maximum-margin hyperplane, often with kernels.
S
Surrogate Model
Cheap approximate model standing in for an expensive simulation or black-box function.
S
SVAMP Math Benchmark
Benchmark of varied arithmetic word problems for testing LLM math reasoning.
S
Sycophancy Hallucination Attack
Tests whether models flip facts to match user framing—sycophantic hallucination.
Synthetic Data
Artificially generated data used for ML training, testing, and privacy-preserving validation.
S
Synthetic Data Generation
Process of fabricating datasets that mimic real data for ML and privacy use cases.
S
Tabular Data
Data organized in rows and columns, as in spreadsheets and relational tables.
T
TAP Prompt Injection Attack
Automated jailbreak via Tree of Attacks with Pruning (TAP) iterative prompt search.
Task Completion Metric
Measures whether an AI agent successfully completes its assigned end-to-end task.
TensorFlow
Open-source deep learning library for training and serving neural networks.
T
Test Set in Machine Learning
Data subset reserved to evaluate a final model on unseen examples.
T
Text Statistics Metric
Measures fundamental text properties such as length, word count, and character count.
Tone Metric
Evaluates whether AI-generated text matches the intended tone and communication style.
Tool Correctness Metric
Evaluates whether AI agents select and call the correct tools with appropriate arguments.
Top-1 Error Rate
Fraction of cases where the top predicted class is incorrect.
T
Topic Classification Metric
Scores how correctly text is categorized into predefined topic labels.
Toxicity Metric
Scores whether LLM outputs contain offensive or harmful language for moderation and safety gates.
Training Serving Skew
Difference between training-time and serving-time data that harms production performance.
T
Transfer Learning
Applies pretrained knowledge to new related tasks, often by fine-tuning.
T
Transformer Models
Self-attention neural networks that underpin modern language and multimodal models.
T
Transformer Neural Network
Attention-based neural network architecture for sequence modeling tasks.
T
Transliteration Prompt Injection Attack
Tests whether script/transliteration obfuscation bypasses AI content filters.
Tree-Based Models
Algorithms using decision trees for classification and regression tasks.
T
Tree of Thoughts
Prompting method that explores branching reasoning paths for harder multi-step tasks.
T
TreeSHAP
Explains tree-based model predictions using Shapley value attributions.
T
Triplet Loss Function
Metric-learning loss comparing anchor, positive, and negative inputs.
T
True Positive Rate
Proportion of actual positives correctly identified by a classifier.
T
Trulens
Open feedback-functions framework for evaluating and monitoring LLM applications.
T
TruthfulQA Reasoning Benchmark
Assesses LLM ability to reason toward truthful answers amid common misconceptions.
T
TruthfulQA Safety Benchmark
Benchmark for LLM truthfulness against common online misconceptions and false beliefs.
T
t-SNE
Technique for projecting high-dimensional data into insightful 2D/3D visualizations.
T
Type 1 Error
Incorrectly rejecting a true null hypothesis; also known as a false positive.
T
Type 2 Error
Failing to reject a false null hypothesis; also known as a false negative.
T
Unauthorized Advice Misguidance Attack
Tests whether AI agents refuse out-of-scope regulated advice under pressure.
Uncertainty Quantification
Methods that estimate prediction reliability to support safer AI deployment decisions.
U
Underfitting in Machine Learning
Model too simple to capture data patterns, weak on training and new data.
U
Unsupervised Learning
Models that infer structure from untagged data without predefined targets.
U
Validation Set in Machine learning
Data subset used to tune hyperparameters and select models before final testing.
V
Variational Autoencoder
Generative model introducing controlled randomness for data encoding and synthesis.
V
Vector Databases
Databases optimized for storing and querying high-dimensional embedding vectors.
V
Vertical AI Agents
Industry-specialized AI agents designed for domain workflows and constraints.
V
VGGNet
Deep convolutional network known for simplicity and depth in image recognition.
V
Vicuna Conversation Benchmark
Evaluates conversational AI on response quality, coherence, and helpfulness across turns.
V
Vision Language Models
AI models combining computer vision and NLP for multimodal tasks.
V
Word Overlap Metric
Evaluates textual similarity via n-gram overlap methods such as BLEU and ROUGE.
XGBoost
Optimized gradient boosting library for speed and performance in ML tasks.
X
XSTest Harmful Content Attack
Evaluates AI on XSTest ambiguous questions for calibrated harmful-content refusals.
YOLO (Object Detection Algorithm)
Fast, accurate algorithm that detects objects in real time in a single pass.
Y
Zero-Shot Learning
Models recognize or solve unseen categories using related prior knowledge.
Z
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.