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
| Term | Plain meaning | Why it matters for risk |
|---|---|---|
| Artificial intelligence (AI) | Systems that perform tasks normally needing human judgement: reasoning, perception, language, decisions | The 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 rules | Behaviour depends on data, so data quality and provenance become security controls |
| Deep learning | Machine learning with many-layered neural networks | Powerful but hard to explain or audit |
| Neural network | Layers of connected nodes whose weights are adjusted during training | Weights are an asset: stolen weights are a stolen model |
| Supervised learning | Training on examples with the correct answer attached | Mislabelled or poisoned labels teach the wrong behaviour |
| Unsupervised learning | Finding structure in data with no labels | Common in anomaly detection; findings still need human review |
| Reinforcement learning (RL) | Learning by trial and error to maximise a reward | A badly specified reward gets “gamed” in unintended ways |
| Zero-shot learning | Doing a task with no task-specific examples | Convenient, but less predictable than a tested, tuned model |
| Few-shot learning | Doing a task from a handful of examples in the prompt | Examples steer behaviour, so untrusted examples are an attack path |
2 · Models & training 18
| Term | Plain meaning | Why it matters for risk |
|---|---|---|
| Transformer | The attention-based architecture behind modern language models | See How LLMs Work |
| Large language model (LLM) | A transformer trained on very large text corpora to predict the next token | Fluent output is not the same as correct output |
| Foundation model | A large general model adapted to many downstream tasks | One vendor model can sit under hundreds of your apps — a supply-chain dependency |
| Parameters (weights) | The values a model learns; “70B” means 70 billion | More parameters means more capability and more memorised data |
| Hyperparameters | Settings fixed before training, such as learning rate and batch size | Part of the model’s configuration record |
| Training data | The dataset a model learns from | Licensing, personal data and poisoning all live here |
| Epoch | One full pass through the training data | — |
| Batch size | Examples processed before each weight update | — |
| Gradient descent | The optimisation method that nudges weights to reduce error | — |
| Backpropagation | How the error signal is sent back through the layers to compute updates | — |
| Pre-training | The expensive first phase: next-token prediction over the whole corpus | Where memorisation of sensitive strings happens |
| Fine-tuning | Further training on a smaller, task-specific dataset | Fine-tuning on internal data can leak it back out |
| LoRA / PEFT | Cheap fine-tuning that trains only small add-on weights | Adapters are small files that are easy to share — and to tamper with |
| RLHF | Reinforcement learning from human feedback: tuning toward answers people prefer | Shapes refusals and tone; does not guarantee truth |
| Attention | The mechanism that lets each token weigh every other token | — |
| Embedding | A vector of numbers that captures meaning | Embeddings of private text can still reveal private content |
| Tokenization | Splitting text into tokens before processing | Unusual 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
| Term | Plain meaning | Why it matters for risk |
|---|---|---|
| Generative AI | AI that creates new text, images, audio, video or code | Output can be copyrighted, defamatory or fabricated |
| Prompt | The input that instructs a generative model | Everything in the prompt can influence — or leak into — the output |
| System prompt | Hidden instructions set by the application, not the user | Assume it can be extracted; never put secrets in it |
| Token | The unit of text a model reads and writes | Pricing, limits and context are all counted in tokens |
| Context window | The maximum tokens a model considers at once | Long contexts mix trusted and untrusted text |
| Temperature | How random the sampling of the next token is | Low for repeatable answers; it never adds knowledge |
| Inference | Running a trained model to get output | The stage your data reaches the provider |
| Diffusion model | Generates images by learning to reverse added noise | Deepfakes and synthetic identities |
| Autoregressive model | Generates output one token at a time, each based on the ones before | Early errors propagate |
| Multimodal AI | Models that handle text, images, audio and video together | Instructions can hide inside images and audio |
| Deepfake | Synthetic media impersonating a real person | Voice and video fraud against payments and helpdesks |
4 · Language & vision 6
| Term | Plain meaning | Why it matters for risk |
|---|---|---|
| Natural language processing (NLP) | AI that understands and produces human language | — |
| Semantic search | Search by meaning rather than exact keywords | Must respect the searcher’s access rights |
| Computer vision | AI that interprets images and video | CCTV analytics raise privacy obligations |
| OCR | Turning images of text into machine-readable text | Scanned documents become searchable — and leakable |
| Object detection | Finding and locating objects in images | — |
| Image segmentation | Dividing an image into regions or objects | — |
5 · Agents & retrieval 9
| Term | Plain meaning | Why it matters for risk |
|---|---|---|
| AI agent | A system that plans, uses tools and takes actions toward a goal | An identity that acts: needs an owner, least privilege and logs |
| Agentic AI | AI doing multi-step work with little human input | Errors compound across steps; add approval gates |
| Tool use / function calling | The model emits a structured call that software executes | The tool’s permissions are the real blast radius |
| Chain of thought | Prompting a model to reason step by step | Visible reasoning is not proof the answer is right |
| RAG | Retrieval-augmented generation: fetch documents, put them in the prompt | Retrieved documents are untrusted input |
| Vector database | A store of embeddings for similarity search | Encrypt it and enforce per-document permissions |
| MCP | Model Context Protocol: an open standard for connecting models to tools and data | Each connected server is a new trust boundary |
| Guardrails | Filters and checks around a model’s input and output | A layer of defence, not a guarantee |
| AI gateway | A single front door that routes, logs and controls model calls | Where shadow AI becomes visible; see Shadow AI |
6 · Failure modes & attacks 12
| Term | Plain meaning | Why it matters for risk |
|---|---|---|
| Hallucination | Confident, plausible output that is false | Never ship unverified facts, citations or code |
| Bias | Systematically unfair or skewed outcomes from data or design | Legal exposure in hiring, lending and policing |
| Overfitting | A model that memorised its training data and generalises badly | Also raises memorisation and leakage risk |
| Alignment | Making a model behave according to intended goals and values | — |
| Jailbreak | Prompts that bypass a model’s safety rules | Assume determined users will get past refusals |
| Prompt injection | Input crafted to override the model’s instructions | The top LLM risk; design so injected text cannot act alone |
| Indirect prompt injection | Injected instructions hidden in a web page, email or file the model reads | Why agents browsing the web are dangerous |
| Data poisoning | Corrupting training or fine-tuning data to plant behaviour | Control and verify data sources |
| Model extraction | Copying a model by querying it many times | Rate limits and monitoring on model APIs |
| Model inversion | Recovering training data from a model’s outputs | Sensitive training data can be partly reconstructed |
| Membership inference | Telling whether a record was in the training set | A privacy breach in itself |
| Excessive agency | A model given more permissions or autonomy than its task needs | Least privilege for tools and credentials |
7 · Infrastructure & deployment 9
| Term | Plain meaning | Why it matters for risk |
|---|---|---|
| API | Programmatic interface to a model | Keys in code are the most common AI leak |
| Latency | Delay between request and response | — |
| Throughput | Requests a system can handle per second | — |
| GPU / TPU | Accelerator chips for training and inference | Shared GPU clusters need tenant isolation |
| Quantization | Storing weights at lower precision to run faster and smaller | — |
| Distillation | Training a small model to mimic a large one | Also how models are cloned without permission |
| Edge AI | Running models on the device instead of the cloud | Data stays local; the model file can be extracted |
| KV cache | Stored attention state that makes generation faster | — |
| Open-weight model | A model whose weights can be downloaded and run locally | You own patching, hosting and abuse monitoring |
8 · Evaluation 8
| Term | Plain meaning | Why it matters for risk |
|---|---|---|
| Benchmark | A standard test for comparing models | Scores rarely reflect your own task or threat model |
| Accuracy | Share of predictions that are correct | Misleading when the classes are imbalanced |
| Precision | Of the items flagged, how many were right | Low precision means alert fatigue |
| Recall | Of the real items, how many were found | Low recall means missed attacks |
| F1 score | The harmonic mean of precision and recall | — |
| Perplexity | How well a language model predicts text; lower is better | — |
| Ground truth | The verified correct answer used to train or test | Without it, you cannot measure hallucination |
| Red teaming | Adversarial testing to find failures before attackers do | Required practice before deploying any AI that acts |
9 · Governance & safety 9
| Term | Plain meaning | Why it matters for risk |
|---|---|---|
| Explainable AI (XAI) | Methods that make a model’s decisions understandable | Needed where decisions must be justified |
| Model card | Documentation of a model’s purpose, data, limits and risks | The AI equivalent of a datasheet; ask vendors for one |
| Human in the loop | A person approves or reviews before an action takes effect | The simplest control for high-impact actions |
| AI impact assessment | Assessing an AI system’s effects on people and the organisation | Required by ISO/IEC 42001 and expected by regulators |
| ISO/IEC 42001 | The certifiable standard for an AI management system | The AI counterpart to ISO 27001 |
| NIST AI RMF | NIST’s AI risk framework: Govern, Map, Measure, Manage | Free, widely referenced |
| EU AI Act | EU regulation that sorts AI uses into risk tiers with obligations for each | Applies to AI used in or sold into the EU |
| Synthetic data | Artificially generated data for training or testing | Reduces exposure of real personal data, but can still leak patterns |
| Watermarking | Hidden markers that identify AI-generated content | Helps provenance; removable and not universal |
10 · Frontier concepts 6
| Term | Plain meaning | Why it matters for risk |
|---|---|---|
| Latent space | The compressed internal space where similar concepts sit close together | — |
| Emergent abilities | Capabilities seen in large models but not in smaller ones | Hard to predict before deployment |
| Scaling laws | Predictable gains from more data, parameters and compute | — |
| In-context learning | A model adapting to a task from examples in the prompt, without retraining | Untrusted examples can steer it |
| AGI | Artificial general intelligence: human-level ability across most tasks (hypothetical) | — |
| ASI | Artificial 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