Resources
Glossary
Plain-language definitions of the terms used across VERITRAIL — the method, evidence labels, runtime guardrails, and the regulations and standards it maps to.
38 terms
- Adversarial testingRuntime guardrails and AI security
Deliberately attacking an AI system — prompt injection, jailbreaks, data leakage — before and after release to find weaknesses while they can still be fixed.
Also: red teaming, red-teaming
Related controls: IS-14
- AI agentRuntime guardrails and AI security
An AI system that plans and takes actions through tools — updating records, sending messages, issuing refunds — rather than only answering questions.
Also: AI agents, agentic AI, agentic workflows
- AI inventoryVERITRAIL method
The owned, classified register of every AI system, model, agent and vendor AI feature an organisation uses, with its owner, purpose, data and risk tier.
Also: AI system inventory, AI system of record
- AI management systemRegulation and standards
The policies, roles, processes and controls an organisation uses to govern AI, as specified by ISO/IEC 42001.
Also: AIMS
- Conformity assessmentRegulation and standards
The process of showing that a high-risk AI system meets EU AI Act requirements before it is placed on the market. VERITRAIL helps organise evidence but does not perform one.
- ControlVERITRAIL method
A requirement in the VERITRAIL Control Library, with an objective, an accountable owner role, examples of evidence and indicative mappings to NIST AI RMF, ISO/IEC 42001 and the EU AI Act.
Also: controls
- Critical gapEvidence and assurance
A gap with material privacy, security, safety, legal or operational exposure that needs priority action.
Also: critical gaps
- Default-denyRuntime guardrails and AI security
An agent may call only the tools and operations its policy lists; anything unlisted is denied, escalated or sent for approval, never silently allowed.
Also: default deny
Related controls: IS-11
- DeployerRegulation and standards
Under the EU AI Act, the organisation that uses an AI system under its authority, as distinct from the provider that develops it or places it on the market. Obligations differ by role.
Also: deployers
- EU AI ActRegulation and standards
The European Union's AI regulation (Regulation (EU) 2024/1689). It sets risk-based obligations for providers and deployers of AI systems, including prohibited practices and requirements for high-risk systems. VERITRAIL's mappings to it are indicative.
Also: Regulation (EU) 2024/1689
- Evidence labelEvidence and assurance
The status on every VERITRAIL finding — Verified, User-asserted, Inferred, Missing evidence or Critical gap — so readers can see what is known and what is assumed.
Also: evidence labels, evidence status
- Evidence packEvidence and assurance
The records that show a control operated for an AI system — approvals, test results, monitoring and logs — assembled on request for an auditor, customer or regulator.
Also: evidence packs, audit pack
- Gap registerEvidence and assurance
A list of governance gaps in which every line records what the gap is, how sure we are (its evidence label), how serious it is, who owns it and when it must be fixed.
Also: gap registers
- General-purpose AI modelRegulation and standards
An EU AI Act term for a model trained on broad data that can perform a wide range of tasks and be built into many systems; its providers have their own obligations.
Also: general-purpose AI models, GPAI
- Governance BaselineVERITRAIL method
A fixed-scope 4–6 week VERITRAIL engagement that validates a self-assessment with evidence and produces a gap register and a remediation roadmap.
Also: VERITRAIL Governance Baseline
- High-risk AI systemRegulation and standards
An EU AI Act category (Article 6 and Annex III) covering uses such as recruitment, credit scoring and access to essential services, with requirements for risk management, data governance, documentation and human oversight. It is not the same as a VERITRAIL High risk tier.
Also: high-risk AI systems
- Human approvalRuntime guardrails and AI security
A runtime decision that pauses a high-impact action — a deletion, external send, payment, export or anything over a threshold — until a designated approver accepts it. The approver's identity is recorded.
Also: REQUIRE_HUMAN_APPROVAL, human approval for high-impact actions
Related controls: IS-13
- Human oversightRuntime guardrails and AI security
Measures that let people understand, monitor and intervene in an AI system's operation, including overriding or stopping it.
- InferredEvidence and assurance
Concluded from answers or metadata rather than from evidence, such as a suggested risk tier or a likely gap.
- ISO/IEC 42001Regulation and standards
The international standard for an AI management system, published in 2023. Organisations can be certified against it by accredited bodies; VERITRAIL does not certify conformity.
- JailbreakRuntime guardrails and AI security
An attempt to make a model produce content or take actions its safeguards are meant to prevent.
Also: jailbreaks
Related controls: IS-14
- Kill switchRuntime guardrails and AI security
A tested way to disable an agent, tool, model or tenant at once. In VERITRAIL Shield, Contain mode denies every request in the policy's scope.
Also: kill switches, Contain mode
Related controls: IS-15
- Maturity levelVERITRAIL method
The band a 0–100 maturity score falls in: Initial (0–39), Developing (40–59), Defined (60–74), Managed (75–89) or Optimized (90–100). Self-assessment scores are indicative and labelled User-asserted.
Also: maturity levels
- Missing evidenceEvidence and assurance
Evidence the control requires has not been provided.
- NIST AI RMFRegulation and standards
The US National Institute of Standards and Technology's voluntary AI Risk Management Framework (AI RMF 1.0, 2023), organised into four functions: GOVERN, MAP, MEASURE and MANAGE.
Also: NIST AI Risk Management Framework
- PillarVERITRAIL method
One of the five groups that organise the VERITRAIL controls: Accountability & Operating Model (AO), Enterprise AI Inventory (EI), Governance & Lifecycle Control (GL), Integrity, Safety & Security (IS), and Stewardship, Evidence & Assurance (SE).
Also: pillars
- Policy-as-codeRuntime guardrails and AI security
Runtime rules written as versioned configuration instead of prose, so they can be reviewed, tested, compared, approved and rolled back like code.
Also: policy as code
- Prohibited-reviewVERITRAIL method
The highest VERITRAIL risk tier. It means "stop and get legal review" because the use case may fall into a prohibited practice; it is not a finding that the use case is prohibited.
Related controls: EI-02
- Prompt injectionRuntime guardrails and AI security
Instructions hidden in user input or retrieved content that try to override an AI system's intended behaviour, for example telling an agent to ignore its policy or reveal data.
Also: prompt-injection
- RedactionRuntime guardrails and AI security
Masking personal or confidential data in prompts, responses, tool parameters and tool outputs before it is stored or passed on.
Also: sensitive-data redaction
Related controls: IS-04
- Release gateVERITRAIL method
A checkpoint an AI system passes before go-live or a material change, scaled to its risk tier, with the approver and any conditions recorded.
Also: release gates, approval gate, approval gates
- Risk tierVERITRAIL method
VERITRAIL's classification of an AI use case as Prohibited-review, High, Medium or Low. The tier sets how much assessment, approval and monitoring the use case needs. A suggested tier is labelled Inferred until an accountable approver confirms it, and it is not the EU AI Act's legal classification.
Also: risk tiers, risk tiering
Related controls: EI-02
- Rollout modeRuntime guardrails and AI security
The stage a runtime policy runs in: Observe (log only), Warn (allow and warn), Protect (redact and transform), Enforce (all decisions apply), with Contain as the kill switch.
Also: rollout modes, staged rollout
- Tamper-evident logEvidence and assurance
A log in which each entry includes a hash of the one before it, so any later edit or deletion can be detected.
Also: tamper-evident logs, tamper-evident decision log, hash-chained log
Related controls: SE-08
- TILLERVERITRAIL method
The assessment and assurance studio in the VERITRAIL workspace. Its six pillars spell TILLER — Trust, Inventory, Lifecycle, Lawfulness, Engineering and Runtime — and each maps to a NIST AI RMF function.
Also: TILLER assessment
- Tool manifestRuntime guardrails and AI security
The declared list of tools, APIs and data stores an agent can reach, and the operations it can perform on each.
Also: tool manifests
Related controls: EI-07
- User-assertedEvidence and assurance
The owner says a control is in place, but no evidence has been reviewed yet. Self-assessment answers carry this label.
- VerifiedEvidence and assurance
Evidence exists and a named reviewer has examined it.