AI Glossary
55 terms used across the courses, each defined in plain language and linked to the lessons that explain it.
Core AI vocabulary
AI system that pursues a goal with autonomy, takes actions via tools, and iterates based on outcomes — distinguished from a single-prompt…
The broader shift from AI that answers questions to AI that does work — systems that plan, call tools, check their own results, and carry…
The mechanism at the heart of the transformer: for every token, the model weighs how relevant every other token in the context is, lettin…
Systematic unfairness in model behavior, usually inherited from patterns in training data. Shows up as skewed outputs for certain groups,…
Maximum number of tokens a model can process in one call (input + output). Modern frontier models offer 200K to 2M+ tokens.
Synthetic audio, image, or video that convincingly imitates a real person. Powered by the same generative models used for legitimate medi…
Training a smaller, cheaper model to imitate a larger one's outputs. The main route by which frontier-level capability trickles down into…
Numerical vector representation of content where semantically similar items cluster in high-dimensional space. Foundation of vector searc…
A structured test suite for AI behavior — benchmarks, graded tasks, or automated judges that measure whether a model or agent actually do…
Further training a base model on domain examples to alter its behavior. Compared to prompting: more expensive, slower to update, behavior…
AI that produces new content — text, images, audio, video, code — rather than only classifying or ranking existing content. LLMs and diff…
The layer of controls wrapped around a model: input screening, output filtering, blocked topics, structured formats, and approval gates. …
Fluent, confident model output not grounded in reality. Mitigated through grounding (RAG), citations, constrained outputs, and human-in-t…
Design pattern where humans approve or oversee AI actions, especially at consequential or irreversible decision points.
Running a trained model to get an answer — as opposed to training, which builds the model. Training happens rarely and costs a lot once; …
Neural network trained on massive text corpora to predict tokens. The substrate of modern generative AI.
Open standard (originated at Anthropic, broadly adopted) for AI hosts to discover and call tools, resources, and prompts via a uniform in…
A short, structured document describing a model: intended use, training data summary, evaluation results, limitations, and known failure …
A model that works across more than one medium — reading images, hearing audio, or generating video as well as text. Frontier models are …
A model whose trained weights are published for anyone to download, run, and fine-tune (Llama, DeepSeek, Qwen, Gemma families). Contraste…
One learned number inside a neural network. Model "size" is the parameter count — billions to trillions of these values, tuned during tra…
Everything you send a model in one request: the instruction, the context, the examples, the data. Prompt quality is the cheapest, fastest…
Malicious instructions embedded in inputs (often via retrieved data) that override intended model behavior. Indirect prompt injection is …
Pattern of retrieving relevant content from a private knowledge base via vector search and feeding it into the LLM prompt at query time. …
A model trained (typically with reinforcement learning) to think before answering — generating internal chains of reasoning and spending …
The standing instructions an application gives the model before any user input — persona, rules, tools, and boundaries. Users rarely see …
A sampling setting that controls how adventurous a model's word choices are. Low temperature: focused and repeatable. High temperature: v…
Unit of input or output processed by an AI model. For text, typically 3-4 characters of English. Images are tokenized as patches, video a…
Chopping raw text (or images, or audio) into the discrete units a model actually processes. Explains many odd model behaviors — why count…
The process of building a model: showing it enormous amounts of data and adjusting its parameters until its predictions improve. Followed…
The neural network architecture behind modern AI, built on the attention mechanism. Introduced in 2017; nearly every frontier language mo…
Security & agent-platform terms
Open protocol for agent-to-agent discovery and communication, championed by Google and now stewarded under the Linux Foundation. Companio…
Google's open-source framework for building AI agents in code. Agents built with it can run anywhere, including on Agent Engine, the mana…
The enforcement point for agent traffic in Google's Gemini Enterprise Agent Platform. It checks each agent-to-tool and agent-to-agent con…
Google's central catalog of an organisation's agents and MCP servers in Gemini Enterprise Agent Platform. Agent Gateway uses it to decide…
AI Bill of Materials — inventory of models, training data, agents, and MCP connections in an AI system. Emerging compliance and security …
Email-based fraud where an attacker impersonates an executive, vendor, or trusted party to trick the target into wiring funds, sharing cr…
Security tooling focused on the endpoint — process activity, file behavior, command-line forensics. Heavy users of classical ML for behav…
Google Cloud's platform for building, deploying, governing and optimising AI agents, announced at Google Cloud Next on 22 April 2026 as t…
Google Security Operations: Google's cloud-native SIEM and SOAR platform, formerly Chronicle and Siemplify. In 2026 it added AI agents fo…
Unified threat intelligence offering combining Mandiant TI and VirusTotal capabilities.
Security category focused on detecting attacks against identity systems — credential theft, session hijacking, MFA fatigue, lateral movem…
Frontline-incident-response firm acquired by Google; the source of Mandiant Consulting, Managed Defense, and the bulk of Google threat in…
Public knowledge base of adversary tactics and techniques targeting AI systems — the AI counterpart to MITRE ATT&CK. The neutral referenc…
Google Cloud's screening service for AI prompts and responses, detecting prompt injection, jailbreaks, sensitive-data leaks and harmful c…
Identities for service accounts, API keys, machines. Distinguished from agent identities, which are autonomous and goal-oriented.
The industry-standard list of the most critical risks in LLM-powered applications — prompt injection, excessive agency, supply chain, dat…
Google Cloud's cloud security platform (CNAPP): posture management, threat detection and vulnerability management. Its AI Protection feat…
Centralized platform for collecting, normalizing, and analyzing security telemetry. Well-known examples include Splunk, Microsoft Sentine…
The playbook-and-automation layer that orchestrates response actions across security tools. Increasingly combined with AI agents that dec…
Secure Production Identity Framework For Everyone — open standard for workload identity, increasingly used as the basis for giving AI age…
Google SecOps normalized schema for security telemetry. Search, detection rules, and analytics operate over UDM.
Detection technique that builds behavioral baselines for users and machine entities, then flags deviations. Predominantly unsupervised ML…
Security platform category that unifies detection across endpoint, network, identity, and cloud. The "extended" generally means EDR plus …
Detection rule language used by Google SecOps. Now supports multi-stage queries for sophisticated detection logic.