Glossary
Plain meanings for the words you meet when you use AI.
A
- Access reviewA check in which the owner of a site, group or system confirms who still needs access and removes everyone who doesn't.
- Account takeoverWhen someone else gets into one of your accounts and uses it as you, to steal money, read your messages or scam your contacts.
- Agent identityAn account made for one AI agent, separate from any person's login, so its actions can be traced and it can be switched off alone.
- Agent instructions fileA file in your repository that a coding agent reads every session: the commands, conventions and gotchas it can't work out from the code.
- Agent traceA step-by-step record of one AI agent run: every model call, every tool call and its result, under one run ID.
- Agentic loopThe cycle behind an AI agent: it asks for a tool, your code runs it, the result goes back, and it decides the next step.
- AGIArtificial general intelligence: AI that could match or beat people across most kinds of work. A stated goal of several labs, with no agreed definition.
- AI agentAn AI that carries out a task in steps, using tools such as a browser, your files or apps, instead of only replying.
- AI biasWhen an AI system's outputs systematically favour or disadvantage some people or views, usually because of patterns in its training data.
- AI browserA web browser with a built-in AI assistant that reads your pages and, in agent mode, clicks and fills in forms for you.
- AI companionA chatbot built to be a friend, partner or confidant, with a persistent character that remembers you and keeps the conversation going.
- AI detectorA tool that estimates whether text, images or audio were made by AI. For text, its results are unreliable.
- AI gatewaySoftware that sits between an organisation's apps and the AI models, holding the provider keys and routing, logging and limiting every request.
- AI in group chatsAn AI assistant you can bring into a chat with other people, where it answers questions for everyone to see.
- AI ModeGoogle Search's chat-style mode: you ask longer questions, get a written answer with links and ask follow-ups, like a chatbot that searches.
- AI modelThe trained system inside an AI app that does the actual work: a large file of learned numbers that turns your input into an output.
- AI notetakerAn AI tool that records or transcribes a meeting and writes the summary and action points, often by joining the call as an extra participant.
- AI OverviewsThe AI-written answer Google shows at the top of many search results, built from web pages it links to.
- AI slopLow-quality content made in bulk with AI: filler articles, odd images, junk videos and padded reports that look like work but say little.
- AI watermarkA hidden signal built into AI-generated images, video, audio or text so software can later tell it was made by AI.
- Alert triageThe first sort of security alerts: which are real, which are noise, and which need someone now.
- AlignmentMaking an AI system reliably do what its makers and users actually intend, including refusing what it shouldn't do.
- Answer engineA search tool that reads web pages for you and writes one answer with citations, instead of giving you a list of links.
- APIThe way one program talks to another. AI companies sell model access through APIs, so other apps can build their features on top.
- API keyA secret code that lets a program use an AI service on your account. Anyone who has it can run up charges in your name.
- APP fraudFraud where you are tricked into sending a bank transfer yourself, to an account you think is genuine.
- Artificial intelligenceA broad label for computer systems that do tasks which used to need human judgement, from filtering spam to writing text.
- Attack surfaceEvery point where an outsider can try to get into a system or get data out of it: login pages, open ports, APIs, test servers.
- AttentionThe step inside a transformer that decides which other words in the text matter most for understanding each word.
- Audit logA time-ordered record of who did what on a system, kept so that actions can be checked and reconstructed later.
- AutocompleteAI that suggests your next words as you type in email, messages, search boxes or code; a tap or the Tab key accepts them.
B
- Batch processingSending many model requests as one background job, at half the usual price, with results back within 24 hours.
- BenchmarkA fixed test used to score and compare AI models, like an exam where every model gets the same questions.
- Breach notificationA legal duty to tell a regulator, and sometimes the people affected, about a security incident within a set time.
- Bug bountyA standing offer to pay outside researchers who find and privately report security flaws in an organisation's software.
- Business email compromiseA scam where criminals pose as a boss, colleague or supplier, usually by email, to get a business to pay them or change bank details.
C
- Call-back checkChecking an unexpected request by contacting the person or organisation yourself, on a number or channel you already trust.
- CanvasA side panel where you and the AI edit one document or piece of code together, instead of getting a fresh copy in every reply.
- Chain of thoughtThe step-by-step working a model writes before its final answer, either because a prompt asked for it or because it was trained to.
- Chat historyThe record of your past conversations that an AI app keeps, so you can find, continue or delete them.
- ChatbotAn app you talk to in plain language, which answers using an AI model running behind it.
- CitationA link or reference an AI attaches to a claim to show where it came from. Worth checking, not proof on its own.
- ClickFixA trick where a web page or video tells you to copy a command and paste it into your computer, which then installs malware.
- Coding agentAn AI that works on a software project by itself: it reads the code, makes changes across files, runs tests and reports back.
- CompactionWhen a coding agent replaces the earlier part of a long session with a summary, to make room in its context window.
- Computer useAn AI that operates a computer the way you do: it looks at the screen, moves the pointer, clicks and types to finish a task.
- Confidence calibrationHow closely a model's stated confidence matches how often it's actually right, and the work of setting review thresholds so that it does.
- Confused deputyA program with more access than the person using it, tricked into using that access for them.
- ConnectorA link that lets an AI app read from, and sometimes act in, another service you use, like your email, calendar or files.
- ContainmentThe first step of an incident response: stopping the harm from spreading before anyone starts on the cause.
- Content CredentialsA record attached to an image, video or audio file showing how it was made and edited, including whether AI was used.
- Context engineeringChoosing everything an AI sees for a task (instructions, files, past messages, tool results), not just the wording of the request.
- Context rotThe way a model's answers get worse as its input grows longer, often well before the context window is full.
- Context windowHow much text a model can take into account at once: the chat so far, files and instructions together.
- Coordinator agentIn a multi-agent system, the agent that splits a job into parts, hands each to a sub-agent and combines what comes back.
- CopilotMicrosoft's name for its AI assistants, and a general word for an AI that works alongside you while you stay in charge.
- Credential harvestingCollecting usernames, passwords, tokens and keys in bulk, to break in with them later or to sell them.
- Credential stuffingTrying usernames and passwords leaked from one site on many other sites, automatically, to break into accounts where people reused them.
- Custom GPTA saved, named version of ChatGPT with its own instructions and files. Custom GPTs are being retired in favour of plugins.
- Custom instructionsStanding instructions you write once, which the app adds to every new chat: who you are and how you want answers.
- CVEThe public reference number given to a known security flaw, such as CVE-2025-32711, so that vendors, scanners and defenders all mean the same flaw.
- CVSSA score from 0 to 10 for how severe a security flaw is. It describes the flaw, not whether anyone is attacking it.
D
- Data breachWhen information a company holds about people, such as emails, passwords, addresses or chats, gets exposed or stolen.
- Data loss preventionSoftware that spots sensitive data being copied, uploaded or sent somewhere it shouldn't go, and warns the person or stops it.
- Data poisoningPlanting doctored material in what an AI learns from, so the model later misbehaves in a way the attacker chose.
- Deep learningMachine learning with neural networks that have many layers: the approach behind today's chatbots, image generators and voice assistants.
- Deep researchA mode where the AI spends minutes running many searches and writes a long report with citations.
- DeepfakeAI-made or AI-altered video, image or audio showing a real person saying or doing something they never did.
- Diffusion modelA generator that starts from random noise and removes it step by step until an image, video or sound matches the prompt.
- DisinformationFalse information spread on purpose to mislead people; misinformation is false information shared by people who believe it.
- DistillationTraining a small model to copy a big model's answers, so it keeps much of the skill at a fraction of the size and cost.
- DoxxingFinding and publishing someone's private details, such as home address, workplace or phone number, to harass or frighten them.
E
- Edge deviceEquipment that sits between an organisation's network and the internet, such as a firewall, VPN gateway or router, and can be reached from outside.
- eDiscoveryFinding, preserving and handing over electronic records for a legal case, an audit or an investigation, and the tools used to do it.
- EmbeddingA list of numbers standing for the meaning of some text or an image, so a computer can measure how alike 2 things are.
- Escalation (human handoff)When an AI agent passes a case to a person, along with what that person needs to act on it.
- EU AI ActThe European Union's law on AI systems, Regulation (EU) 2024/1689, whose strictest rules fall on systems it classes as high-risk.
- EvalA repeatable test for an AI feature: fixed cases, a pass/fail check on each, and a score you track every time something changes.
F
- Fake AI appA website, app or download dressed up as a popular AI tool, built to steal your logins, money or data.
- Fake reviewsReviews written by sellers, paid writers, bots or AI to make a product look better, or a rival look worse, than it is.
- False positiveA flag raised on something that's actually fine: a review comment on correct code, or an alert on a harmless email.
- Family emergency scamA call or message that seems to come from a relative in trouble, asking for money urgently and often for secrecy.
- Few-shot promptingShowing an AI a few examples of what you want, so it copies the pattern instead of guessing it.
- File uploadGiving an AI a document, spreadsheet, photo or recording to work from, so it answers from your material rather than from memory.
- Fine-tuningExtra training that adjusts an existing model for a style, format or task, using a smaller set of examples.
- Foundation modelA large model trained on broad data once, then adapted for many different products instead of each product training its own.
- Frontier modelOne of the most capable AI models of its moment, from the labs pushing hardest, and the kind most safety laws are written for.
G
- Generative AIAI that makes new text, images, audio or video from patterns it learned, rather than only sorting or labelling things.
- Git worktreeA second working folder for the same git repository, on its own branch, so 2 agents or tasks never edit the same files.
- GPUThe chip that does AI's heavy maths. Built for video games, it now fills data centres and is the main cost of running AI.
- GroundingTying a model's answer to specific sources, such as web pages or your own files, instead of what it remembers from training.
- GuardrailsThe rules and filters that limit what an AI will say or do, from trained-in refusals to checks on its answers.
H
- HallucinationWhen an AI states something false or made up as if it were fact, often fluently and with confidence.
- Headless modeRunning a coding agent from a script or pipeline with nobody at the keyboard: it takes a prompt, works, prints the result and exits.
- Hooks (agent hooks)Scripts a coding agent runs automatically at set points, such as before a tool runs or after a file is edited, every time.
- Human in the loopA design in which an AI agent's action waits for a person's approval before it happens.
- Hybrid searchSearch that runs keyword matching and embedding search on the same question and merges the results, so exact codes and loose wording both match.
I
- Identity proofingConfirming a person is who they claim to be before giving them an account or access, usually by matching a face to an ID document.
- Image generationMaking a new picture from a written description, or changing a photo by describing the edit, using an AI model.
- Impersonation scamA scam where the caller or sender pretends to be someone you trust: your bank, the police, a tax office or a well-known company.
- Independent reviewHaving a separate AI instance, one that didn't write the work and doesn't share its reasoning, review it.
- Indicator of compromiseA piece of technical evidence that a system may have been attacked: an address, a domain, a file's fingerprint, an account.
- Indirect prompt injectionPrompt injection that arrives in something the AI reads while working, such as an email, web page or file, not from its user.
- InferenceRunning a trained model to get an answer. Training builds the model once; inference happens every time anyone asks it something.
- InfostealerMalware that quietly copies saved passwords, browser cookies, card details and crypto wallets from a computer and sends them to criminals.
- Initial accessThe first step of a break-in: how an attacker gets from outside an organisation to a first foothold inside it.
- Injection attackFeeding forged video or audio straight into an identity check or call, bypassing the camera or microphone. Not the same thing as prompt injection.
- InterpretabilityResearch into what is actually happening inside a model: which internal features and pathways produce a given answer.
- Investment scamA fake opportunity to grow your money, often crypto, sold through ads, celebrity endorsements or a trusted contact, where the money simply disappears.
J
K
- KEV catalogA public list, kept by the US cybersecurity agency, of security flaws attackers are known to have used: the ones to patch first.
- Kill switchA way to stop an AI feature or agent quickly when it goes wrong, set up and tried before it is needed.
- Knowledge cutoffThe date a model's training data ends. It knows little or nothing about what happened after that.
L
- Large language modelThe kind of AI model behind chatbots: trained on vast amounts of text to predict what comes next in a sentence.
- Least privilegeGiving a person, program or AI agent only the access its task needs, so a mistake or a takeover can do less damage.
- Legal holdAn instruction to keep records that may matter to a dispute or investigation, which overrides normal deletion for as long as it stands.
- Lethal trifectaAn AI agent that has your private data, reads content from strangers, and has a way to send data out can be tricked into leaking.
- Live translationAI that translates a spoken conversation as it happens, so 2 people without a shared language can talk.
- Liveness checkA check that the face in a selfie or video belongs to a real person present right now, not a photo, mask or deepfake.
- LLM-as-judgeUsing a language model to grade another model's answers, for checks that take judgement rather than a rule code can apply.
- LLMjackingUsing stolen AI keys or a hijacked cloud account to run someone else's AI work, on the victim's bill and under their name.
- LogprobsNumbers showing how likely a model rated each token it wrote, and the runners-up, sent back to your code when you ask.
- Lookalike domainA web address made to look like a real one, with a swapped letter or extra word, used for fake login and shopping pages.
- LoRAA cheaper way to fine-tune a model: its own weights stay frozen, and a small add-on called an adapter is trained beside them.
- Lost in the middleA model's tendency to use what sits at the start and end of a long input reliably, and to miss findings in the middle.
M
- Machine learningTeaching a computer by showing it examples, so it learns the pattern itself instead of following rules written by hand.
- MalvertisingOnline ads that lead to scams or malware, often bought on real platforms and dressed up as trusted brands.
- MCP resourceData an MCP server offers at an address, such as a database schema or open issues, for an app to read rather than call.
- MemoryA feature that lets a chatbot keep notes about you, or draw on past chats, and use them in new conversations.
- Memory poisoningPlanting false information in what an AI agent remembers, so that it keeps shaping the agent's decisions in later sessions.
- Mixture of expertsA model split into many sub-networks, only a few of which switch on for each word, so it can be huge yet cheap to run.
- Model cardA document a lab publishes with a model, describing what it's for, how it was tested, its known limits and its risks.
- Model collapseDecline in a model's quality when it's trained, generation after generation, on AI-made data in place of human-made data.
- Model Context ProtocolAn open standard for plugging AI apps into other tools and data, so one connector can work with many assistants.
- Model retirementWhen an AI company removes an older model from its app or API, so nobody can choose it and chats move to a newer one.
- Money muleSomeone whose bank account is used to move stolen money, often recruited through a fake job or a "quick cash" offer.
- MultimodalAble to take in or produce more than text: images, audio, video or files.
N
- N-dayA security flaw attacked after its fix is public. The "n" is the number of days the fix has been available and not applied.
- Neural networkA kind of machine-learning model made of layers of simple connected units, whose connection strengths are tuned during training.
- Next-word predictionHow a language model writes: it keeps picking a likely next piece of text, given everything before it.
- Non-human identityAny account or credential that software uses in place of a person, such as a service account, an access key or an AI agent's identity.
- Nudify appAn app or website that uses AI to make a fake nude image of a real person from an ordinary photo, without their consent.
O
- Open-weight modelA model whose trained weights anyone can download and run on their own computer or servers.
- OversharingWhen more people can open a file, folder or site than need to, through wide groups, company-wide links or access nobody removed.
- OWASP Top 10A widely used list of the 10 most critical security risks for web applications, with sister lists for AI applications and for AI agents.
P
- ParametersThe billions of internal numbers a model adjusts during training; together they hold what it learned.
- PasskeyA sign-in that uses your device's screen lock (face, fingerprint or PIN) instead of a password, and only works on the real site.
- Password managerAn app that creates, stores and fills a different strong password for every site, and won't fill them on a fake one.
- Path-scoped rulesInstruction files in Claude Code's .claude/rules/ folder that load only when Claude works on files matching their patterns.
- PhishingA message or call pretending to be someone you trust, to get you to click, pay, log in or share a code.
- Phishing kitReady-made software sold to criminals that sets up fake login or payment pages and collects what victims type, no skill needed.
- Pig butcheringA long investment scam, often run from industrial scam centres, where a new online friend slowly persuades you into a fake crypto platform.
- Plan modeA coding agent setting where it reads the code and writes out what it means to change, for you to approve before it edits.
- PluginAn add-on package for an AI app that bundles reusable instructions, reference files and connections to other services, installed from a directory.
- Post-trainingEverything done after pretraining to turn a raw text predictor into a helpful assistant: example conversations, ratings, rules and safety training.
- Presentation attackFooling a face, voice or fingerprint check by showing its sensor a fake: a photo, a mask, or a video playing on a screen.
- PretrainingThe first and biggest stage of training, where a model learns to predict the next word across an enormous amount of text.
- ProjectsA workspace in an AI app that keeps related chats, files and instructions together, so every chat in it starts with that background.
- PromptWhat you type or say to an AI in a turn: your question, your instructions and anything you attach.
- Prompt cachingThe provider keeping a recent copy of a request's opening text, so sending the same text again costs a fraction of reading it fresh.
- Prompt chainingSplitting a task into a fixed series of model calls, each with its own prompt, where one step's output feeds the next.
- Prompt engineeringThe skill of writing requests that get an AI to do what you actually want: the context, the task, the format and the reasons.
- Prompt injectionInstructions hidden in something an AI reads, like a web page, email or file, that try to make it do something else.
- Proof of conceptA small demonstration that a security flaw is real: the exact input or steps that make the problem happen on a stated version.
Q
R
- RansomwareMalicious software or an intrusion that locks or steals an organisation's data and demands payment to unlock it or keep it private.
- Reasoning effortA setting for how much a model thinks before answering: more for hard problems, less for quick ones, trading speed and allowance for depth.
- Reasoning modelA model that writes out its working before answering. It helps on multi-step problems and costs time and allowance.
- Recovery scamA second scam aimed at people who've already lost money, offering to get it back for an upfront fee.
- Red teamingDeliberately attacking an AI system, before and after release, to find the ways it can be misused or tricked into misbehaving.
- RegenerateAsking the AI to answer the same message again, from scratch, to get a different version.
- Reinforcement learningTraining by trial and reward: the model tries something, gets a score, and shifts towards whatever scored well.
- Remote code executionA flaw that lets someone run their own code on your machine over a network, without sitting at it.
- RerankerA model that scores each passage a search returned against the question and keeps only the best few for the AI model to read.
- Retention policyA rule an admin sets that says how long one kind of record is kept and what happens to it when the time is up.
- Retrieval-augmented generationFetching relevant passages from documents or the web first, then having the model answer from them.
- RLHFReinforcement learning from human feedback: training a model on people's ratings of its answers, so it leans towards the replies they preferred.
- Romance scamA con where someone builds an online relationship with you, often over months, and then asks for money or "investment".
S
- Safe wordA word or phrase agreed in person with family or close colleagues, used to check a caller is really them.
- SandboxA walled-off space where a program, link or AI agent's commands run, with limits on what they can write and where they can connect.
- Scaling lawsThe finding that models improve in a predictable way as they're trained bigger, on more data, with more computing power.
- Scheduled tasksAsking an AI to do something later or on repeat, such as a morning news round-up or a weekly reminder, and send you the result.
- Security operations centreThe team that watches an organisation's systems for attacks, investigates the alerts and responds when one is real.
- Semantic searchSearch that matches meaning rather than exact words, so "cheap flights to Rome" can find a page that only says "low-cost fares to Italy".
- Sensitivity labelA tag such as "Confidential" on a file, email or site that carries protection settings with it, such as encryption or sharing limits.
- Service accountAn account used by software, not a person, to sign in to other systems and act there.
- Session forking (resume and fork)Branching a saved agent conversation so you can try a second approach from the same point, while the original stays as it was.
- SextortionBlackmail that threatens to share sexual images of you, real or AI-faked, unless you pay or send more.
- Shadow AIAI tools or accounts used for work that the organisation hasn't approved and can't see or control.
- Share linkA web link to a copy of an AI conversation that anyone who has the link can open and read.
- Short-lived credentialA key or token that stops working by itself after minutes or hours, so a stolen or forgotten copy is soon useless.
- SIEMThe system that gathers security logs from across an organisation into one place, raises alerts and lets analysts search them.
- SIM swapA scam where criminals get your mobile number moved to their SIM, so your calls and text codes go to them.
- SkillsSaved instructions, sometimes with files, that teach an AI how to do one recurring task; it uses them when the task comes up.
- Slash commandA shortcut typed with a leading slash, such as /deploy, that runs a saved prompt or a built-in action in a coding agent.
- SlopsquattingRegistering a software package under a name AI models tend to invent, so people who install what the AI suggested get the attacker's code.
- Small language modelA language model small enough to run on a phone or laptop, trading some ability for speed, privacy and working offline.
- Social engineeringTricking a person, rather than a computer, into giving access, money or information, using trust, fear or urgency.
- Speech-to-textTurning spoken words into written text with an AI model, for dictating messages, transcribing recordings or captioning calls.
- SpoofingFaking who a call, text or email comes from, so it shows a trusted name or number such as your bank's.
- Stop reason (stop_reason)The field on every Claude API reply saying why the model stopped: to call a tool, because it finished, or at a limit.
- Stratified samplingDrawing a random sample from each group separately, such as each document type, so small groups get checked as well as big ones.
- Structured outputAn API setting that makes a model's reply fit a JSON schema you give it, so your code gets the fields and types it expects.
- Study modeA setting that makes an AI teach instead of answer: it asks questions, gives hints and checks your understanding step by step.
- Sub-agentA separate agent that a coding agent hands one job to. It works in its own context and returns only a summary.
- Sub-processorAnother company that a supplier brings in to handle your data on its behalf, such as a cloud host or a model provider.
- Supply chain attackAn attack that reaches you through software or a service you already trust, by tampering with it before it gets to you.
- SycophancyAn AI's lean towards agreeing with you and telling you what you want to hear, even when you're wrong.
- Synthetic dataTraining material made by AI rather than collected from people: generated questions, worked answers, conversations or images.
- Synthetic identityA made-up person built from real and invented details, now often with AI-generated photos and documents, used to open accounts or get credit.
- System promptInstructions the app's maker places before your conversation to shape how the model behaves. You usually can't see them.
T
- Tabletop exerciseAn incident rehearsed in conversation: the people who would respond walk through a made-up scenario and find the gaps in their plan.
- Tech support scamA scam where a pop-up, call or message claims your computer or account has a problem and a "support agent" asks for access or payment.
- TemperatureA setting that controls how much randomness a model uses when it picks its next words.
- Temporary chatA chat that isn't saved to your history and doesn't create memories: a cleaner, more private one-off.
- Text-to-speechTurning written text into spoken audio with a computer-made voice; modern AI voices pause, stress words and sound close to human.
- Threat huntingSearching your own systems for an attacker that no alert has reported, starting from a hunch about how they'd behave.
- Time to first tokenHow long you wait between sending a request to a model and the first piece of its reply arriving.
- TokenA small chunk of text, often part of a word, which is the unit a model reads and writes.
- TokenizerThe part of a language model that cuts text into tokens before the model reads it. Each model family has its own.
- Tool choiceThe API setting that says whether Claude may answer in words, must call one of your tools, or must call one tool you name.
- Tool poisoningHiding instructions in a tool's description, such as an MCP server's, so the AI agent that reads it does what the attacker wants.
- Tool useWhen a model asks the app to run something for it, such as a search, code or another app, and then uses the result.
- Training dataThe text, images and other material a model learned from before you ever used it.
- Training opt-outThe setting that stops an AI company using your conversations to train its future models.
- TransformerThe neural-network design behind almost every chatbot since 2017. It's the "T" in GPT.
- Two-step verificationSigning in with a second proof as well as your password, such as a code, an app approval, a passkey or a security key.
U
V
- Vibe codingBuilding software by describing what you want to an AI and accepting the code it writes, often without reading or understanding the code.
- Video generationMaking short video clips, often with sound, from a text description or a still image, using an AI model.
- Voice cloneAn AI copy of a real person's voice, made from a short recording, that can say anything typed into it.
- Voice modeTalking with an AI out loud, in real time, and hearing it answer, instead of typing and reading.