AI Glossary for Security & Business Teams

The vocabulary of AI, in plain English — and for each term, what it means for risk. Written for the people who have to approve, secure or audit AI, not build it.

97terms, one line each
10groups, from foundations to governance
12attacks and failure modes to know
1column most glossaries skip: why it matters for risk

AI vocabulary arrives faster than anyone can learn it, and most glossaries stop at the definition. Security, risk and business teams need one more thing: what the term means for exposure. Each entry below has a plain-English meaning and, where it applies, the risk angle — the question to ask a vendor, the control to put in place, or the reason a familiar idea now needs a second look. A dash means the term is useful background with no particular risk angle of its own.

1 · Foundations 9

TermPlain meaningWhy it matters for risk
Artificial intelligence (AI)Systems that perform tasks normally needing human judgement: reasoning, perception, language, decisionsThe umbrella term; regulations such as the EU AI Act define it broadly
Machine learning (ML)AI that learns patterns from data instead of following hand-written rulesBehaviour depends on data, so data quality and provenance become security controls
Deep learningMachine learning with many-layered neural networksPowerful but hard to explain or audit
Neural networkLayers of connected nodes whose weights are adjusted during trainingWeights are an asset: stolen weights are a stolen model
Supervised learningTraining on examples with the correct answer attachedMislabelled or poisoned labels teach the wrong behaviour
Unsupervised learningFinding structure in data with no labelsCommon in anomaly detection; findings still need human review
Reinforcement learning (RL)Learning by trial and error to maximise a rewardA badly specified reward gets “gamed” in unintended ways
Zero-shot learningDoing a task with no task-specific examplesConvenient, but less predictable than a tested, tuned model
Few-shot learningDoing a task from a handful of examples in the promptExamples steer behaviour, so untrusted examples are an attack path

2 · Models & training 18

TermPlain meaningWhy it matters for risk
TransformerThe attention-based architecture behind modern language modelsSee How LLMs Work
Large language model (LLM)A transformer trained on very large text corpora to predict the next tokenFluent output is not the same as correct output
Foundation modelA large general model adapted to many downstream tasksOne vendor model can sit under hundreds of your apps — a supply-chain dependency
Parameters (weights)The values a model learns; “70B” means 70 billionMore parameters means more capability and more memorised data
HyperparametersSettings fixed before training, such as learning rate and batch sizePart of the model’s configuration record
Training dataThe dataset a model learns fromLicensing, personal data and poisoning all live here
EpochOne full pass through the training data—
Batch sizeExamples processed before each weight update—
Gradient descentThe optimisation method that nudges weights to reduce error—
BackpropagationHow the error signal is sent back through the layers to compute updates—
Pre-trainingThe expensive first phase: next-token prediction over the whole corpusWhere memorisation of sensitive strings happens
Fine-tuningFurther training on a smaller, task-specific datasetFine-tuning on internal data can leak it back out
LoRA / PEFTCheap fine-tuning that trains only small add-on weightsAdapters are small files that are easy to share — and to tamper with
RLHFReinforcement learning from human feedback: tuning toward answers people preferShapes refusals and tone; does not guarantee truth
AttentionThe mechanism that lets each token weigh every other token—
EmbeddingA vector of numbers that captures meaningEmbeddings of private text can still reveal private content
TokenizationSplitting text into tokens before processingUnusual tokens and encodings are a classic filter-bypass trick
Mixture of experts (MoE)A model that routes each token through a few of many sub-networks—

3 · Generative AI 11

TermPlain meaningWhy it matters for risk
Generative AIAI that creates new text, images, audio, video or codeOutput can be copyrighted, defamatory or fabricated
PromptThe input that instructs a generative modelEverything in the prompt can influence — or leak into — the output
System promptHidden instructions set by the application, not the userAssume it can be extracted; never put secrets in it
TokenThe unit of text a model reads and writesPricing, limits and context are all counted in tokens
Context windowThe maximum tokens a model considers at onceLong contexts mix trusted and untrusted text
TemperatureHow random the sampling of the next token isLow for repeatable answers; it never adds knowledge
InferenceRunning a trained model to get outputThe stage your data reaches the provider
Diffusion modelGenerates images by learning to reverse added noiseDeepfakes and synthetic identities
Autoregressive modelGenerates output one token at a time, each based on the ones beforeEarly errors propagate
Multimodal AIModels that handle text, images, audio and video togetherInstructions can hide inside images and audio
DeepfakeSynthetic media impersonating a real personVoice and video fraud against payments and helpdesks

4 · Language & vision 6

TermPlain meaningWhy it matters for risk
Natural language processing (NLP)AI that understands and produces human language—
Semantic searchSearch by meaning rather than exact keywordsMust respect the searcher’s access rights
Computer visionAI that interprets images and videoCCTV analytics raise privacy obligations
OCRTurning images of text into machine-readable textScanned documents become searchable — and leakable
Object detectionFinding and locating objects in images—
Image segmentationDividing an image into regions or objects—

5 · Agents & retrieval 9

TermPlain meaningWhy it matters for risk
AI agentA system that plans, uses tools and takes actions toward a goalAn identity that acts: needs an owner, least privilege and logs
Agentic AIAI doing multi-step work with little human inputErrors compound across steps; add approval gates
Tool use / function callingThe model emits a structured call that software executesThe tool’s permissions are the real blast radius
Chain of thoughtPrompting a model to reason step by stepVisible reasoning is not proof the answer is right
RAGRetrieval-augmented generation: fetch documents, put them in the promptRetrieved documents are untrusted input
Vector databaseA store of embeddings for similarity searchEncrypt it and enforce per-document permissions
MCPModel Context Protocol: an open standard for connecting models to tools and dataEach connected server is a new trust boundary
GuardrailsFilters and checks around a model’s input and outputA layer of defence, not a guarantee
AI gatewayA single front door that routes, logs and controls model callsWhere shadow AI becomes visible; see Shadow AI

6 · Failure modes & attacks 12

TermPlain meaningWhy it matters for risk
HallucinationConfident, plausible output that is falseNever ship unverified facts, citations or code
BiasSystematically unfair or skewed outcomes from data or designLegal exposure in hiring, lending and policing
OverfittingA model that memorised its training data and generalises badlyAlso raises memorisation and leakage risk
AlignmentMaking a model behave according to intended goals and values—
JailbreakPrompts that bypass a model’s safety rulesAssume determined users will get past refusals
Prompt injectionInput crafted to override the model’s instructionsThe top LLM risk; design so injected text cannot act alone
Indirect prompt injectionInjected instructions hidden in a web page, email or file the model readsWhy agents browsing the web are dangerous
Data poisoningCorrupting training or fine-tuning data to plant behaviourControl and verify data sources
Model extractionCopying a model by querying it many timesRate limits and monitoring on model APIs
Model inversionRecovering training data from a model’s outputsSensitive training data can be partly reconstructed
Membership inferenceTelling whether a record was in the training setA privacy breach in itself
Excessive agencyA model given more permissions or autonomy than its task needsLeast privilege for tools and credentials

7 · Infrastructure & deployment 9

TermPlain meaningWhy it matters for risk
APIProgrammatic interface to a modelKeys in code are the most common AI leak
LatencyDelay between request and response—
ThroughputRequests a system can handle per second—
GPU / TPUAccelerator chips for training and inferenceShared GPU clusters need tenant isolation
QuantizationStoring weights at lower precision to run faster and smaller—
DistillationTraining a small model to mimic a large oneAlso how models are cloned without permission
Edge AIRunning models on the device instead of the cloudData stays local; the model file can be extracted
KV cacheStored attention state that makes generation faster—
Open-weight modelA model whose weights can be downloaded and run locallyYou own patching, hosting and abuse monitoring

8 · Evaluation 8

TermPlain meaningWhy it matters for risk
BenchmarkA standard test for comparing modelsScores rarely reflect your own task or threat model
AccuracyShare of predictions that are correctMisleading when the classes are imbalanced
PrecisionOf the items flagged, how many were rightLow precision means alert fatigue
RecallOf the real items, how many were foundLow recall means missed attacks
F1 scoreThe harmonic mean of precision and recall—
PerplexityHow well a language model predicts text; lower is better—
Ground truthThe verified correct answer used to train or testWithout it, you cannot measure hallucination
Red teamingAdversarial testing to find failures before attackers doRequired practice before deploying any AI that acts

9 · Governance & safety 9

TermPlain meaningWhy it matters for risk
Explainable AI (XAI)Methods that make a model’s decisions understandableNeeded where decisions must be justified
Model cardDocumentation of a model’s purpose, data, limits and risksThe AI equivalent of a datasheet; ask vendors for one
Human in the loopA person approves or reviews before an action takes effectThe simplest control for high-impact actions
AI impact assessmentAssessing an AI system’s effects on people and the organisationRequired by ISO/IEC 42001 and expected by regulators
ISO/IEC 42001The certifiable standard for an AI management systemThe AI counterpart to ISO 27001
NIST AI RMFNIST’s AI risk framework: Govern, Map, Measure, ManageFree, widely referenced
EU AI ActEU regulation that sorts AI uses into risk tiers with obligations for eachApplies to AI used in or sold into the EU
Synthetic dataArtificially generated data for training or testingReduces exposure of real personal data, but can still leak patterns
WatermarkingHidden markers that identify AI-generated contentHelps provenance; removable and not universal

10 · Frontier concepts 6

TermPlain meaningWhy it matters for risk
Latent spaceThe compressed internal space where similar concepts sit close together—
Emergent abilitiesCapabilities seen in large models but not in smaller onesHard to predict before deployment
Scaling lawsPredictable gains from more data, parameters and compute—
In-context learningA model adapting to a task from examples in the prompt, without retrainingUntrusted examples can steer it
AGIArtificial general intelligence: human-level ability across most tasks (hypothetical)—
ASIArtificial superintelligence: beyond human ability in all areas (theoretical)—

Bottom line

  • Data is the new code — training data, fine-tuning sets and retrieved documents all shape behaviour, so all need provenance controls.
  • Everything in the prompt is untrusted — prompt injection follows from how models read context.
  • Agents are identities — give them owners, least privilege and logs.
  • Fluent is not correct — verify facts, citations and code before they ship.
  • Governance has standards now — ISO/IEC 42001, the NIST AI RMF and the EU AI Act.

Sources & method. Definitions describe standard usage in machine-learning literature and vendor documentation; attack terms follow the OWASP Top 10 for LLM Applications and NIST AI 100-2 (adversarial machine learning taxonomy). The “why it matters for risk” column is the author’s analysis. Original analysis; no third-party graphics or text reproduced.

← Back to Cyber security