Resources · Glossary

Every term, defined once.

54 terms, alphabetical, each one to three sentences and cross-referenced to the chapter that develops it. Where a term has a canonical spelling in the style guide, that spelling is used here and everywhere in the book.

A

Agent (agentic AI): An AI system that acts — browses, executes code, calls APIs, moves data or delegates to other agents — under delegated authority, rather than only producing text. Agents are the hardest object to govern because their behaviour is emergent and their actions have external effects. ch. 01
Agent registry: The runtime-aware inventory of every non-human actor — model, service and agent — each with an owner, a declared scope, a status and a kill switch, fed by a runtime data path rather than typed by hand. It is the artefact that answers "what AI is running?". ch. 04 ch. 06
AESIA: Spain's Agencia Española de Supervisión de la Inteligencia Artificial, based in A Coruña — the first operational national AI supervisor in the EU, with sanctioning powers since 2025. ch. 08
AI Act (EU): Regulation (EU) 2024/1689, the EU's horizontal, risk-tiered law for AI, amended by the Digital Omnibus. It classifies systems by risk (prohibited, high-risk, limited, minimal) and imposes obligations accordingly. ch. 08
AI governance engineer: The person who holds the capability of AI governance engineering and is accountable for the three questions in production; a capability and a role, not necessarily a job title. ch. 06
AI governance engineering: The application of engineering practice — systems thinking, product thinking and code — to the governance of AI systems; measured by realised risk reduction and audit-ready evidence. ch. 01
AI Office: The European Commission body that supervises general-purpose AI and coordinates AI Act enforcement, with investigation powers and the ability to levy penalties on GPAI providers. ch. 08
AI RMF functions: The four core functions of the NIST AI Risk Management Framework — Govern, Map, Measure, Manage — used throughout the book as a mapping target for controls. ch. 08
AIBOM: AI bill of materials: the machine-readable inventory of an AI system's components — models, datasets, dependencies — in formats such as CycloneDX ML-BOM or the SPDX 3.0 AI profile. ch. 04 ch. 05
AICM: The CSA AI Controls Matrix, a control framework (v1.1, 247 control objectives across 18 domains) that maps to ISO 42001, ISO 27001 and NIST AI RMF and underpins STAR for AI. ch. 08
AIMA: The OWASP AI Maturity Assessment, reported at v1.0 (Aug 2025), which scores the breadth of an AI security and governance programme across domains. ch. 07
AIMS: An AI management system: the governance structure, roles, controls and continual-improvement loop that ISO/IEC 42001 certifies. An AIMS is not the AI Act's Article 17 quality management system. ch. 07 ch. 08
Annex III: The AI Act annex listing high-risk use cases (biometrics, critical infrastructure, employment, essential services, law enforcement, migration, justice); obligations for these phase in under the Digital Omnibus timeline. ch. 08
ASI01–ASI10: The ten risks of the OWASP Top 10 for Agentic Applications 2026 — from ASI01 Agent Goal Hijack and ASI02 Tool Misuse through ASI03 Agent Identity & Privilege Abuse to ASI10 Rogue Agents — the canonical threat list for agents. ch. 05
ATLAS: MITRE's Adversarial Threat Landscape for Artificial-Intelligence Systems, a knowledge base of adversary tactics and techniques against AI, including agent-specific techniques. ch. 05 ch. 10
Audit-ready evidence: Evidence emitted as a by-product of the build in a form an auditor can read directly — machine-readable, signed, timestamped — so the audit is a query, not a collection project. ch. 01 ch. 04

C

CIMD: Client ID Metadata Document: the MCP mechanism (2026 spec) by which a client identifies itself via a URL-addressable metadata document, replacing deprecated Dynamic Client Registration. ch. 05
Continuous assurance: Assurance produced continuously from telemetry rather than at a point in time; the control's status is a live query, not an annual sign-off. It is Level 5 of the maturity model. ch. 04 ch. 07

D

Data card: Structured, versioned documentation of a dataset — provenance, lawful basis, rights, composition and known limitations — maintained as code alongside the system. ch. 04
Digital Omnibus: The 2026 reform package amending the EU AI Act (in force 27 Jul 2026), which adjusted the high-risk timeline, added AI Office investigation powers and reworked several articles. ch. 08
DPIA: Data Protection Impact Assessment: the GDPR Article 35 assessment of processing likely to result in high risk to individuals, maintained in this discipline as a versioned artefact, not a one-off document. ch. 04 ch. 05
Drift: The gradual divergence of a model's inputs, outputs or performance from its validated baseline over time; a runtime signal that a control or eval must catch. ch. 04

E

Eval gate: A pipeline stage that fails the build when an eval fails; the mechanism that turns an evaluation into an enforced control rather than a report. ch. 04 ch. 05
Evals: Automated tests of a model's or agent's behaviour — capability, safety and adversarial — run as controls, not as one-off research. ch. 04
Evals as evidence: The principle that the eval run *is* the assurance evidence: a failing eval blocks the build and its structured result is stored as proof the control fired. ch. 03 ch. 04

F

Framework crosswalk: A mapping of one framework's controls onto another's; useful as an index, but a crosswalk proves you read the framework, not that the mapped control fires. ch. 08
FRIA: Fundamental Rights Impact Assessment: the AI Act Article 27 assessment of a high-risk system's impact on rights, maintained here as a versioned, reviewable artefact. ch. 04 ch. 05

G

Governance-as-code: Governance rules expressed as executable code that evaluates pull requests, deployments and runtime calls and returns a decision; the umbrella term of which policy-as-code is the CI/CD subset. ch. 03 ch. 04
GPAI: General-purpose AI: models that can serve many tasks, governed by dedicated AI Act obligations (transparency, copyright, and for systemic-risk models, evaluation and incident reporting) enforced by the AI Office from 2 Aug 2026. ch. 08
Guardian agent: An AI agent whose job is to supervise, check or constrain other agents at runtime; Gartner is reported to project guardian agents as a material share of the agentic market by 2030. ch. 04 ch. 05

H

Harmonised standard: A European standard cited in the Official Journal that, once cited, grants a presumption of conformity with a specific AI Act requirement. As of the book's date, none is yet cited. ch. 08

K

Kill switch: A tested mechanism to stop an agent or system from acting; a precondition of granting autonomy, registered against the agent's identity. ch. 03 ch. 05

M

Machine-readable evidence: Evidence a machine can query, diff and aggregate — OSCAL artefacts, structured eval results, signed logs — as opposed to screenshots and exported spreadsheets. ch. 03 ch. 04
Market surveillance authority: The national authority designated to enforce the AI Act for products placed on its market, with powers to investigate, demand documentation and require corrective action. ch. 08
MCP: Model Context Protocol: an open protocol for connecting AI applications to tools and data sources; its 2026 specification adds OAuth 2.1 resource-server patterns and issuer-bound credentials for agent authorisation. ch. 05
Model card: Structured, versioned documentation of a model — provenance, intended use, capabilities, evaluations and known failure modes — maintained as code. ch. 04
Model risk management: The SR 11-7-tradition practice of validating models for conceptual soundness and back-testing them; a neighbour of this discipline, extended here to runtime behaviour and agents. ch. 01

N

NHI: Non-human identity: the identity of an agent, service account or machine actor. Every NHI gets a registry entry, an owner and a scope before it is allowed to act. ch. 04 ch. 05

O

OSCAL: The Open Security Controls Assessment Language, a NIST machine-readable format for controls, assessments and evidence, used here as the format for audit-ready evidence. ch. 04 ch. 10

P

Paved path: A supported, low-friction default route — a template, library or pipeline — that makes the governed way the easiest way to ship, so engineers adopt governance without asking permission. ch. 03 ch. 06
Policy Card: A JSON-schema, machine-readable governance artefact that declares an agent's allowed and forbidden behaviours for runtime enforcement. ch. 04 ch. 10
Policy-as-code: Governance policy expressed in an executable policy language (OPA/Rego, Cedar) that evaluates in CI/CD and at admission; the narrower, pipeline subset of governance-as-code. ch. 04 ch. 05
Post-market monitoring: The AI Act Article 72 duty to actively monitor a high-risk system's performance and risks after deployment, throughout its lifetime. ch. 08
Presumption of conformity: The legal effect by which conforming to a cited harmonised standard is taken as meeting the corresponding AI Act requirement; unavailable until a standard is OJ-cited. ch. 08

Q

QMS (Art. 17): The quality management system that AI Act Article 17 requires of high-risk providers; distinct from an ISO/IEC 42001 AIMS, which certifies a management system but is not harmonised. ch. 08

R

RAISE Act: New York's Responsible AI Safety and Education Act, a frontier-AI safety law reported to take effect 1 Jan 2027 (status: verify enactment at time of writing; the cited source describes it as awaiting final state action). ch. 08
Realised risk reduction: The measured drop in a named failure mode's rate or blast radius in production — one of the two tests of the discipline, against framework coverage. ch. 01 ch. 03
Red teaming: Structured adversarial testing of a model or agent to elicit failures (jailbreaks, injection, tool misuse) before an attacker does; treated here as an evidence-producing control. ch. 04 ch. 05
Runtime data path: The live connection between production and the governance function — discovery, telemetry and enforcement — without which a registry or dashboard describes the program but cannot see what is running. ch. 04 ch. 07

S

SB 53: California's frontier-AI transparency law (TFAIA), in force 1 Jan 2026, covering large frontier developers training models above 10^26 FLOP, with transparency and safety-framework obligations. ch. 08
Serious incident: Under the AI Act, an incident causing death, serious harm to health, fundamental rights or property, or serious disruption of critical infrastructure, triggering Article 73 reporting duties. ch. 04 ch. 08
Shadow AI: An AI system, model or agent running in production without registering — the failure mode that makes an inventory complete only for the honest. ch. 07
STAR for AI: CSA's security assurance and certification programme for AI, built on the AICM, with a self-assessment tier, an automated "Valid-AI-ted" tier and a Level 2 combining ISO/IEC 42001 with the validated assessment. ch. 07 ch. 08
Systemic risk: Under the AI Act, the risk posed by the most capable general-purpose AI models, triggering extra evaluation, adversarial-testing and incident-reporting duties on their providers. ch. 08