Glossary

AI security terms, defined

One page per term: the definition first, then what it means for securing AI systems and complying with AI rules, and where it appears on this site. Every entry is sourced to the standard, law or organization the term belongs to.

12 terms

A

Agentic AIAgentic AI means systems that use an AI model to plan and act toward a goal with little human input. How Five Eyes agencies, the NCSC and OWASP define it.Also: agentic systems, agentic AI systems, autonomous AI
AI agentAn AI agent is software that repeatedly prompts an AI model and carries out the actions it picks through tools. NIST and OWASP definitions and the controls it needs.Also: LLM agent, autonomous agent, agent hijacking
AI guardrailsAI guardrails are controls that check prompts, retrieved content, outputs and agent actions before they take effect. What they cover, where they fail, what they are not.Also: LLM guardrails, generative AI guardrails, AML.M0020
AI red teamingAI red teaming is structured adversarial testing that finds flaws in AI models and systems before attackers do. Definitions from NIST, CISA, the NCSC and the EU AI Act.Also: AI red-teaming, adversarial testing, LLM red teaming

D

Data poisoningData poisoning is tampering with the data an AI model learns from to plant errors, bias or backdoors. Definitions from NIST, OWASP, MITRE and the EU AI Act.Also: training data poisoning, backdoor poisoning, AML.T0020

G

General-purpose AI model (GPAI)Under the EU AI Act, a general-purpose AI model performs a wide range of tasks and can be built into many systems. The legal test, provider duties and key dates.Also: GPAI, GPAI model, general-purpose AI model with systemic risk

H

High-risk AI systemUnder the EU AI Act, a high-risk AI system is one in a regulated product safety role or an Annex III use. How classification works, the duties and the delayed dates.Also: high-risk AI, Annex III system, Article 6 classification

J

Jailbreak (large language models)A jailbreak is input or tampering that makes a large language model ignore its safety restrictions. How NIST, OWASP and MITRE define it and why it matters.Also: jailbreaking, LLM jailbreak, AML.T0054

M

P

Prompt injectionPrompt injection is input that makes a large language model act against its developer's instructions. What it is, why it persists and where it shows up.Also: direct prompt injection, indirect prompt injection, LLM01:2025

S

Shadow AIShadow AI is staff use of AI tools outside the systems an organization has approved. The NCSC's definition, the data risks it creates and how teams govern it.Also: unsanctioned AI, unapproved AI tools, shadow generative AI

A term missing, or a definition that reads wrong? Request a correction. Longer answers live in the explainers.