AI Security Glossary and Taxonomy

Purpose: Provide common terminology for AI security governance, engineering, red teaming, assurance, vendor evaluation and audit.

Audience: All wiki users.

Term	Definition
AI Usage Governance	Controls that discover, monitor and govern how users interact with AI tools.
AI Data Protection	Controls that detect and prevent leakage of sensitive data through AI interactions.
AI Runtime Security	Controls that protect AI applications during operation.
AI-SPM	AI Security Posture Management; inventory, assess and govern AI assets and configurations.
AI Supply Chain Security	Protection of models, datasets, libraries, plugins, MCP servers and AI dependencies.
Agentic AI Security	Governance of AI agents that can use tools, APIs or workflows.
Prompt Injection	Attack that manipulates model instructions or application behavior through crafted input.
Indirect Prompt Injection	Prompt injection delivered through external or retrieved content.
RAG Security	Security of retrieval-augmented generation systems and their data sources.
Shadow AI	Use of AI tools without approval, governance or security visibility.
MCP	Model Context Protocol; an emerging mechanism for connecting AI systems to tools and data sources.
Tool Calling	AI system invocation of tools, APIs, commands, workflows or plugins.
AI Output Leakage	Sensitive information exposed through generated AI responses.
AI Red Teaming	Adversarial testing of AI systems to validate resilience and controls.
AI Security	The broader discipline of managing security risks associated with AI systems and AI use; the umbrella over security of AI and AI for security.
Security of AI	Protection of AI systems and their lifecycle from malicious or accidental compromise.
AI for Security	Use of AI to perform or augment defensive security functions.
Adversarial Machine Learning	Study and mitigation of attacks that exploit machine-learning behaviour.
Data Poisoning / Model Poisoning	Malicious manipulation of training data or model artifacts.
Excessive Agency	Unnecessary functionality, permissions or autonomy that enables harmful actions (OWASP LLM06).
Policy Enforcement Point	A deterministic component that authorizes or denies an action independently of the model.
Provenance	The origin and transformation history of an artifact or data item.
Robustness	Resilience under relevant perturbations and unexpected conditions.
Evaluation	Systematic measurement against defined criteria.
Model Drift	Changes in behaviour or performance over time, including those caused by data, model, configuration or environment.
LLM	Large Language Model; a model designed to process and generate language.
SLM	Small Language Model; a smaller model optimized for constrained or specialized use cases.
RAG	Retrieval-Augmented Generation; a pattern where the AI system retrieves external content to support responses.
Vector Database	A system used to store and retrieve embeddings or semantic representations.
Agent	AI system capable of planning or acting through tools, APIs or workflows.
Guardrails	Controls that constrain or monitor AI inputs, outputs and behavior.
AI BOM	AI Bill of Materials; inventory of models, datasets, libraries, plugins and dependencies.
Model Registry	Repository for tracking models, versions, metadata and governance status.
AI Assurance	Evidence-based validation that AI systems operate securely, reliably and according to policy.

The distinction between AI security, security of AI and AI for security is explained in Terminology and the Two-Axis Model.

Usage Guidance

Use this glossary to standardize language in risk reviews, control descriptions, architecture documents, vendor questionnaires and audit evidence.