Glossary

The terms that come up again and again in the news, each explained in a few sentences.

AI agent: An AI system that does more than answer: given a goal, it plans steps and carries them out, using tools such as a browser, a code editor or other software.
API: Application programming interface: the agreed way one piece of software asks another for data or a service.
Backdoor: A hidden way into a system that bypasses its normal security, planted by an attacker or, in some cases, built in deliberately.
Benchmark: A standard set of tasks used to measure and compare systems, such as AI models on maths or coding problems, or chips on speed.
CVE: Common Vulnerabilities and Exposures: the public catalogue that gives each known security flaw an identifier such as CVE-2026-12345, so everyone can refer to the same flaw.
Data center: A building full of servers, storage and networking equipment that runs online services and, increasingly, AI models.
Export controls: Government rules on which goods and technologies may be sold abroad, and to whom.
Frontier model: The most capable AI models of their time, usually built by a handful of large labs at great cost.
GPU: Graphics processing unit: a chip built to do many simple calculations at once. Designed for graphics, it turned out to suit AI, and it is now the main hardware for training and running models.
HBM: High-bandwidth memory: memory chips stacked on top of each other and placed beside a processor, so data reaches it much faster.
Inference: Running a trained AI model to get an answer, as opposed to training it.
IPO: Initial public offering: the first time a private company sells shares on a stock exchange, after which anyone can buy them.
Jailbreak: In AI, a prompt written to get a model past its safety rules, so that it produces what it was built to refuse.
Large language model (LLM): A model trained on very large amounts of text to predict what comes next, which lets it write, summarise, translate and answer questions.
Malware: Any software made to harm a computer or its user: to steal data, spy, take control or disrupt it. Viruses, spyware and ransomware are all kinds of malware.
Multimodal: Describes an AI model that works with more than one kind of input or output, such as text, images, audio and video, rather than text alone.
On-device AI: AI that runs on the phone, laptop or other device itself instead of on a company's servers.
Open source: Software whose source code is published under a licence that lets anyone read, use, change and share it.
Open-weight model: A model whose trained parameters, its weights, are published so anyone can download, run and modify it on their own hardware.
Phishing: A scam in which an attacker poses as someone trustworthy, by email, text or a fake website, to trick people into handing over passwords, money or access.
Prompt injection: An attack that hides instructions inside content an AI system reads, such as a web page, an email or a document, so the system follows the attacker's instructions instead of its user's.
Ransomware: Malicious software that locks a victim's files, usually by encrypting them, and demands payment to unlock them.
Remote code execution: A flaw that lets an attacker run their own code on someone else's machine over a network, without physical access.
Robotaxi: A taxi service with self-driving cars and no driver at the wheel, run commercially in a growing number of cities.
Supply-chain attack: An attack that reaches its targets through something they trust, such as a software library, an update or a supplier, instead of attacking them directly.
TPU: Tensor processing unit: Google's own chip designed for AI work, used inside Google and rented to others through its cloud as an alternative to Nvidia's GPUs.
Valuation: What a company is judged to be worth. For a private company it is usually set by the price investors pay for shares in a funding round, not by its revenue or profit.
Zero-day: A security flaw that attackers know about, and often already use, before the software's maker has released a fix: the maker has had zero days to respond.