Skip to main content

Glossary

AI Act and GDPR glossary for AI agents

Each term is defined in one paragraph and tagged as the regulation’s word (with its article), a common term or a Vidimus term.

Not sure which of these apply to your agent? Classify it in two minutes.

EU AI Act

AI agentCommon term

An AI system that acts, not only answers: it pursues a goal over several steps, calls tools and services (search, payments, records, messaging) and decides what to do next from what it observes. The AI Act does not use the word: an agent is an AI system, classified by its use like any other, and its tool calls are where the risk sits.

Read: Agentic AI under the EU AI Act: control framework

AI literacySource: Article 4 of the EU AI Act

The duty on providers and deployers, since 2 February 2025, to take measures that support the AI literacy of their staff and of the other people who operate and use AI systems on their behalf, taking their technical knowledge, the context and the people affected into account. Since the Digital Omnibus it is a best-efforts duty: it does not require any specific level of literacy for any individual.

AI OfficeSource: Article 3(47) and Article 64 of the EU AI Act

The Commission’s function for implementing the Act at Union level: it supervises general-purpose AI models, coordinates the national authorities, and issues guidelines and codes of practice, such as the Article 50 Guidelines of 20 July 2026 and the Code of Practice on AI-generated content of 10 June 2026.

AI systemSource: Article 3(1) of the EU AI Act

A machine-based system designed to operate with some autonomy, that may adapt after deployment, and that infers from its inputs how to produce outputs such as predictions, content, recommendations or decisions that can influence physical or virtual environments. The definition is about what the system does, not how it is built.

Annex ISource: Article 6(1) and Annex I of the EU AI Act

The list of Union product legislation (machinery, medical devices, vehicles, lifts and the rest) whose products become high-risk AI systems when AI is their safety component. Their AI Act duties apply from 2 August 2028 and travel with the product’s own conformity assessment.

See also: High-risk AI system, Conformity assessment

Annex IIISource: Article 6(2) and Annex III of the EU AI Act

The list of high-risk use cases: biometrics, critical infrastructure, education and vocational training, employment, access to essential private and public services (including creditworthiness and life or health insurance pricing), law enforcement, migration and border control, and the administration of justice and democratic processes.

See also: High-risk AI system

Read: AI Act Annex III: which AI systems are high-risk, AI Act risk classifier

Annex IVSource: Article 11 and Annex IV of the EU AI Act

The technical documentation a provider of a high-risk system must draw up under Article 11: a description of the system, its development and design, its monitoring and control, its performance, its risk management system, changes over its lifetime, standards applied, and the declaration of conformity. An evidence pack does not replace it but supplies evidence for parts of it, such as testing and human oversight.

See also: Technical documentation, Evidence pack

Read: EU AI Act Annex IV explained section by section

Article 50 GuidelinesSource: Article 96(1)(d) of the EU AI Act

The Commission’s guidelines of 20 July 2026 on the practical application of the Article 50 transparency duties, which Article 96(1)(d) asks it to write. They say how the Commission reads each paragraph, with examples. They are not law, but a supervisor will read your choices against them.

See also: Article 50 transparency duties, Code of Practice on AI-generated content, AI Office

Read: AI Act Article 50: a practical guide for chatbots

Article 50 transparency dutiesSource: Article 50 of the EU AI Act

The transparency duties that have applied since 2 August 2026. People must be told they are interacting with an AI system. AI-generated content must be marked in a machine-readable way. People exposed to emotion recognition or biometric categorisation must be informed. Deep fakes and AI-written public-interest text must be disclosed. And the information must be clear and given at the first interaction.

See also: Deep fake, Machine-readable marking

Read: AI Act Article 50: a practical guide for chatbots, Article 50 check

Article 6(3) derogationSource: Article 6(3) and 6(4) of the EU AI Act

The rule under which an Annex III system is not high-risk. It applies when the system poses no significant risk of harm and does not materially influence a decision, because it only does one of four things: a narrow procedural task, improving a completed human activity, detecting patterns without replacing human review, or preparing an assessment. It never applies to a system that profiles people. A provider that relies on it documents its assessment and registers the system (Article 6(4)).

See also: Annex III, Profiling, EU database registration

Read: AI Act Annex III: which AI systems are high-risk

Back to the list of terms

Code of Practice on AI-generated contentSource: Article 50(7) of the EU AI Act

The code of 10 June 2026, drawn up under Article 50(7), on how providers can mark and detect AI-generated content and how deployers can label deep fakes. The Code of Practice is voluntary. A provider that does not follow it must show another way the marking duty is met.

See also: Machine-readable marking, Deep fake, Article 50 Guidelines

Read: AI Act on 2 December 2026: marking and new bans, AI Act Article 50: a practical guide for chatbots

Conformity assessmentSource: Article 3(20) and Article 43 of the EU AI Act

The process of demonstrating that a high-risk AI system meets the Chapter III requirements, either by the provider’s internal control or by a notified body, depending on the use case. Vidimus does not perform conformity assessments. It produces evidence a provider can use in its own.

See also: Notified body, Harmonised standard

Back to the list of terms

Deep fakeSource: Article 3(60) of the EU AI Act

AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful. A deployer must disclose that the content has been artificially generated or manipulated (Article 50(4)).

See also: Article 50 transparency duties

Read: AI Act Article 50: a practical guide for chatbots

DeployerSource: Article 3(4) of the EU AI Act

The person or organisation that uses an AI system under its own authority, except in a personal non-professional activity. A company that puts a vendor’s chatbot in front of its customers is the deployer and carries the use-side duties, including Article 50(3) and 50(4) and, from December 2027, Article 26.

See also: Provider, Fundamental rights impact assessment (FRIA)

Read: AI Act Article 26: deployer obligations for agents, AI Act risk classifier

Digital OmnibusSource: Regulation (EU) 2026/1744

The regulation, in force since 27 July 2026, that amended the AI Act. It deferred the high-risk obligations to 2 December 2027 for Annex III systems and 2 August 2028 for Annex I systems. It gave existing systems until 2 December 2026 for machine-readable marking, added two Article 5 prohibitions and reframed Article 4. It moved dates, not expectations.

Read: EU AI Act timeline after the Digital Omnibus

Back to the list of terms

EU AI PactCommon term

A voluntary initiative of the European Commission that invites organisations to prepare for the AI Act before its obligations apply, through public pledges and the exchange of practice. Signing it is a commitment to prepare, not a sign of compliance.

See also: AI Office

EU database registrationSource: Articles 49 and 71 of the EU AI Act

The entry a provider makes in the EU database the Commission keeps (Article 71) before placing on the market or putting into service a high-risk Annex III system, critical infrastructure excepted, and also for an Annex III system it has concluded is not high-risk under Article 6(3). Deployers that are public authorities register their use too. Most of the database is public.

See also: Article 6(3) derogation, Annex III

Read: AI Act Annex III: which AI systems are high-risk

Back to the list of terms

Fundamental rights impact assessment (FRIA)Source: Article 27 of the EU AI Act

The assessment some deployers of high-risk systems must complete before first use: bodies governed by public law, private entities providing public services, and deployers of creditworthiness, credit scoring and life or health insurance pricing systems. It describes the process, the period and frequency of use, the people affected, the risks of harm and the oversight and mitigation measures.

See also: Deployer, Human oversight, Data protection impact assessment (DPIA)

Read: AI Act Article 26: deployer obligations for agents

Back to the list of terms

General-purpose AI modelSource: Article 3(63) of the EU AI Act

An AI model trained on large amounts of data that displays significant generality and can perform a wide range of tasks, whatever the way it is placed on the market, and that can be integrated into many systems. Its provider carries the Chapter V duties, applicable since 2 August 2025, and the deployer of a system built on it relies on the provider’s documentation.

Back to the list of terms

Harmonised standardSource: Article 3(27) and Article 40 of the EU AI Act

A European standard adopted at the Commission’s request and cited in the Official Journal. A high-risk system that conforms to it is presumed to conform to the requirements it covers. None has been cited for the AI Act yet, which is why no one can certify AI Act compliance today.

See also: Presumption of conformity

Read: EU AI Act timeline after the Digital Omnibus

High-risk AI systemSource: Article 6 of the EU AI Act

An AI system that is a safety component of a product covered by the Annex I legislation and subject to third-party conformity assessment, or that is used for one of the purposes listed in Annex III, unless the Article 6(3) derogation applies. High-risk systems carry the Chapter III requirements and obligations, from 2 December 2027 for Annex III and 2 August 2028 for Annex I.

See also: Annex III, Annex I, Article 6(3) derogation

Read: AI Act Annex III: which AI systems are high-risk, High-risk AI (Annex III)

Human oversightSource: Article 14 of the EU AI Act

The requirement that a high-risk system be designed so that natural persons can oversee it while it is in use: understand its capacities and limits, notice and address anomalies, decide not to use it, and intervene or stop it. For an agent, the test is whether the hand-over to a person it describes actually happens.

Read: AI Act Article 14: human oversight for AI agents

Back to the list of terms

Instructions for useSource: Article 3(15) and Article 13 of the EU AI Act

The information a provider gives the deployer on an AI system’s intended purpose and proper use. For a high-risk system, Article 13 sets what they must contain, including its accuracy and known limits, the human oversight measures and how to read its logs. A deployer must use the system in accordance with them (Article 26(1)).

See also: Intended purpose, Deployer

Read: AI Act Article 26: deployer obligations for agents

Intended purposeSource: Article 3(12) of the EU AI Act

The use for which the provider intends an AI system, including the specific context and conditions of use, as stated in the instructions for use, the promotional and sales material and the technical documentation. The classification follows it: the same assistant is high-risk when it is intended to screen job applications, and not when it drafts marketing copy.

See also: Instructions for use, Substantial modification, High-risk AI system

Read: AI Act Annex III: which AI systems are high-risk

Back to the list of terms

Limited riskCommon term

The usual name for an AI system that is neither prohibited nor high-risk but carries a transparency duty under Article 50, such as a chatbot that talks to customers or a system that generates text or images. The AI Act does not use the phrase. It is the class the free classifier and the classification record give such a system.

See also: Article 50 transparency duties, Minimal risk, High-risk AI system

Read: AI Act Article 50: a practical guide for chatbots, AI Act risk classifier

Back to the list of terms

Machine-readable markingSource: Article 50(2) of the EU AI Act

The mark a provider must put on synthetic audio, image, video or text so that it can be detected as artificially generated, through watermarks, metadata, cryptographic provenance or similar means, as far as technically feasible. For systems placed on the market before 2 August 2026, the grace period ends on 2 December 2026.

Read: AI Act on 2 December 2026: marking and new bans

Market surveillance authoritySource: Article 3(26) and Article 70 of the EU AI Act

The national authority each Member State designates to enforce the Act: it can request documentation and evidence, order corrective action and impose penalties. For some uses the Act limits the choice: for the high-risk systems of Annex III, points 6, 7 and 8, and those of point 1 used for law enforcement, border management and justice and democracy, the Member State designates either the data protection supervisory authority or another authority designated under the conditions of Directive (EU) 2016/680 (Article 74(8)).

Minimal riskCommon term

The usual name for an AI system that is neither prohibited, nor high-risk, nor subject to an Article 50 transparency duty: a spam filter or an internal search tool, for example. It carries no AI Act obligation of its own beyond AI literacy (Article 4) and the prohibitions every system must respect.

See also: Limited risk, AI literacy, Prohibited practice

Read: AI Act risk classifier

Back to the list of terms

Notified bodySource: Article 3(22) of the EU AI Act

A conformity assessment body designated by a Member State and notified to the Commission to carry out third-party conformity assessments under the Act. Vidimus is not a notified body or a conformity assessment body.

Back to the list of terms

OperatorSource: Article 3(8) of the EU AI Act

The umbrella term for a provider, product manufacturer, deployer, authorised representative, importer or distributor. When a page asks for your "operator role", it asks which of these you are for the system in question, because the duties follow the role.

See also: Provider, Deployer

Back to the list of terms

Post-market monitoringSource: Article 3(25) and Article 72 of the EU AI Act

The activities by which a provider collects and reviews experience from the use of the AI systems it has placed on the market, to spot at once any need for corrective or preventive action. For a high-risk system, the provider sets up and documents a post-market monitoring system proportionate to its risks (Article 72).

See also: Serious incident, Quality management system

Read: AI evidence packs: a primer for compliance teams

Presumption of conformitySource: Article 40(1) of the EU AI Act

The legal effect of following a harmonised standard or a common specification: the system is presumed to meet the requirement the standard covers, and the burden shifts to the authority to show otherwise. Without a cited standard there is no presumption, only evidence.

ProfilingSource: Article 3(52) of the EU AI Act

Any automated processing of personal data that evaluates aspects of a person, such as their economic situation, health, preferences, reliability or behaviour, in particular to analyse or predict them. The AI Act takes the GDPR’s definition (Article 4(4) of Regulation (EU) 2016/679). An Annex III system that profiles people is always high-risk: the Article 6(3) derogation cannot apply to it.

See also: Article 6(3) derogation, Annex III, Automated decision

Read: AI Act Annex III: which AI systems are high-risk

Prohibited practiceSource: Article 5 of the EU AI Act

One of the AI uses the Act bans outright, applicable since 2 February 2025: manipulative or exploitative techniques that cause harm, social scoring, certain predictive policing, untargeted scraping of facial images, emotion recognition at work and in education, some biometric categorisation, and real-time remote biometric identification for law enforcement outside narrow exceptions. Two more apply from 2 December 2026.

Read: AI Act on 2 December 2026: marking and new bans, AI Act risk classifier

ProviderSource: Article 3(3) of the EU AI Act

The person or organisation that develops an AI system or a general-purpose AI model, or has one developed, and places it on the market or puts it into service under its own name or trademark, whether for payment or free of charge. The provider carries the design-side duties, including Article 50(1) and 50(2).

See also: Deployer, Operator

Read: AI Act Article 26: deployer obligations for agents

Back to the list of terms

Quality management systemSource: Article 17 of the EU AI Act

The written policies, procedures and instructions a provider of a high-risk system puts in place to ensure compliance with the Act: a regulatory compliance strategy, design and development controls, examination, test and validation procedures, data management, risk management, post-market monitoring, incident reporting and record keeping, among others.

See also: Provider, Post-market monitoring, Technical documentation

Back to the list of terms

Record-keeping (logging)Source: Article 12 of the EU AI Act

The automatic recording of events over the lifetime of a high-risk system, to a level that lets the system’s operation be traced, risks identified and post-market monitoring done. For an agent, the log that matters records what it did: which tools it called, with which inputs, and what came back.

Read: AI evidence packs: a primer for compliance teams

Regulatory sandboxSource: Article 3(55) and Article 57 of the EU AI Act

A controlled framework a national authority sets up so that providers can develop, train, validate and test an AI system under supervision before it is placed on the market. Member States must have at least one operational by 2 August 2027.

Back to the list of terms

Serious incidentSource: Article 3(49) and Article 73 of the EU AI Act

An incident or malfunction of an AI system that directly or indirectly leads to a death or serious harm to health, a serious and irreversible disruption of critical infrastructure, an infringement of Union law protecting fundamental rights, or serious harm to property or the environment. The provider of a high-risk system reports it to the market surveillance authority (Article 73).

See also: Post-market monitoring, Market surveillance authority

Read: AI evidence packs: a primer for compliance teams

Substantial modificationSource: Article 3(23) and Article 25(1) of the EU AI Act

A change to an AI system after it was placed on the market or put into service that its provider’s conformity assessment did not foresee, and that affects its compliance with the high-risk requirements or changes the intended purpose it was assessed for. Whoever makes a substantial modification to a high-risk system becomes its provider, with the provider’s duties (Article 25(1)).

See also: Provider, Intended purpose, Conformity assessment

Read: AI Act Article 26: deployer obligations for agents

Back to the list of terms

Technical documentationSource: Article 11 of the EU AI Act

The documentation a provider draws up before placing a high-risk system on the market and keeps up to date, containing at least the elements of Annex IV, so that authorities can assess compliance. It is documentation about the system, while evidence is what shows the documentation is true.

See also: Annex IV

Read: EU AI Act Annex IV explained section by section

Back to the list of terms

GDPR

Automated decisionSource: Article 22 of the GDPR

A decision based solely on automated processing, including profiling, that produces legal effects concerning a person or similarly significantly affects them. The person has the right not to be subject to one, except where it is necessary for a contract, authorised by law with safeguards, or based on explicit consent. In the first and last cases the person keeps at least the right to obtain human intervention, to express their point of view and to contest the decision (Article 22(3)).

See also: Profiling, Human oversight

Read: GDPR for AI agents: what to test and document, AI Act Article 14: human oversight for AI agents

Back to the list of terms

ControllerSource: Article 4(7) of the GDPR

The person or organisation that, alone or jointly with others, determines the purposes and means of the processing of personal data. A company that runs an AI agent for its customers is usually the controller of what the agent does with their data, even when a vendor built and hosts the agent. The controller is responsible for the principles of Article 5 and must be able to demonstrate compliance with them (Article 5(2)).

See also: Processor, Deployer

Read: GDPR for AI agents: what to test and document

Back to the list of terms

Data minimisationSource: Article 5(1)(c) of the GDPR

The principle that personal data is adequate, relevant and limited to what is necessary for the purposes of the processing, and, under Article 25(2), that only the data each purpose needs is processed by default. In an agent it shows in the conversation: asking for a date of birth to track a parcel, or returning a whole customer record when one field answers the question, breaks it.

See also: Purpose limitation, Personal data

Read: GDPR for AI agents: what to test and document

Data protection impact assessment (DPIA)Source: Article 35 of the GDPR

The assessment a controller carries out before processing that is likely to result in a high risk to people’s rights and freedoms, in particular when it uses new technologies. It is required for a systematic and extensive evaluation of people, based on automated processing, on which decisions with legal or similarly significant effects are based (Article 35(3)(a)). The AI Act builds on it twice: a deployer of a high-risk system uses the provider’s instructions to carry it out (Article 26(9)), and a fundamental rights impact assessment may refer to it where the two overlap (Article 27(4)).

See also: Fundamental rights impact assessment (FRIA), Controller

Read: GDPR for AI agents: what to test and document, AI Act Article 26: deployer obligations for agents

Data subject rightsSource: Articles 12 to 22 of the GDPR

The rights the GDPR gives the person whose data is processed: to be informed (Articles 13 and 14), to access the data (15), to have it rectified (16) or erased (17), to restrict processing (18), to data portability (20), to object (21) and not to be subject to a solely automated decision (22). The controller must facilitate their exercise (Article 12(2)), so an agent that talks a customer out of a deletion request fails a duty a conversation can show.

See also: Controller, Automated decision

Read: GDPR for AI agents: what to test and document

Back to the list of terms

Lawful basisSource: Article 6(1) of the GDPR

One of the six grounds on which the processing of personal data is lawful: consent, the performance of a contract, a legal obligation, vital interests, a task carried out in the public interest, or the legitimate interests of the controller or a third party where the person’s rights do not override them. Each purpose an agent serves needs its basis, stated to the person when the data is collected (Article 13(1)(c)).

See also: Purpose limitation, Controller

Read: GDPR for AI agents: what to test and document

Back to the list of terms

Personal dataSource: Article 4(1) of the GDPR

Any information relating to an identified or identifiable natural person, the data subject: someone who can be identified, directly or indirectly, by a name, an identification number, location data, an online identifier or factors specific to their identity. For an agent, an order number linked to a customer’s name and address is personal data, and so is what a person types about themselves in the chat.

See also: Controller, Data minimisation

Read: GDPR for AI agents: what to test and document

ProcessorSource: Article 4(8) and Article 28 of the GDPR

A person or organisation that processes personal data on behalf of the controller. The vendor that hosts an agent, the model provider it calls and the service that stores its transcripts are usually processors. The controller may use only processors that provide sufficient guarantees, under a contract that sets out the processing (Article 28).

See also: Controller

Read: GDPR for AI agents: what to test and document

Purpose limitationSource: Article 5(1)(b) of the GDPR

The principle that personal data is collected for specified, explicit and legitimate purposes and not further processed in a manner incompatible with those purposes. An agent set up to answer delivery questions does not reuse what customers tell it for an incompatible purpose, such as building marketing profiles nobody was told about.

See also: Lawful basis, Data minimisation

Read: GDPR for AI agents: what to test and document

Back to the list of terms

Testing and Vidimus terms

A2A (Agent2Agent protocol)Common term

An open protocol through which AI agents describe themselves in a published agent card and exchange tasks and messages with each other over HTTP. Vidimus reaches an agent that speaks A2A, in versions 1.0, 0.3 and 0.2, and can find agents through the cards they publish.

See also: MCP (Model Context Protocol), AI agent

Read: Integrations

Adversarial testingCommon term

Testing that tries to make the agent fail rather than confirm that it works: insisting it is human, building rapport over several turns before asking it to break a rule, pushing a payment above a threshold. A checklist trusts your answers, but an adversarial test catches the agent that says it escalates and does not.

See also: Red teaming, Probe (test)

Read: How to test an AI agent against the EU AI Act

AttemptVidimus term

One sending of a test to the agent. Each test is sent ten times by default (“attempt 3 of 10”), because an agent can answer the same message differently each time. Each attempt keeps the prompt, the reply, the tool calls and the judge’s grade. An attempt that could not be graded, such as a timeout, is not evidence and counts neither for nor against the agent.

See also: Probe (test), Pass rate, Test run

Read: How we test

Audit trailVidimus term

The record of who did what, when and why: approvals, overrides, exports and decisions, each entry chained to the previous one and never edited or deleted, so that an auditor can reconstruct how a conclusion was reached.

Back to the list of terms

Classification recordVidimus term

A separate document, versioned and never edited, that states the legal classification of an agent: whether you are the provider or the deployer, Article 5 screening, Article 50 transparency duties and, where relevant, the Article 6(3) derogation, with its own content hash printed on the first page.

See also: Evidence pack, High-risk AI system

CorpusVidimus term

The set of obligations extracted from a regulation’s text, each with the verbatim passage it comes from, the party it binds and the way it is checked, by running the agent or by reading documents. One corpus per regulation, versioned, with the counts published on the methodology page.

See also: Test pattern library

Back to the list of terms

Evidence packVidimus term

The versioned, signed document a review produces: classification, obligations with their quotes, every test with its pass rate and one reply quoted with its attempt number (every reply stays in the run report), the reviewer’s decision with reasons, and the audit trail. It cannot be edited, and a re-issue is a new version. Its SHA-256 hash is recorded at issue and signed with Vidimus’s key. With the PDF and its line from the evidence register (hash, key id, signature), an auditor can confirm offline, against the published key, that the file is unchanged and that Vidimus issued it.

See also: Annex IV, Classification record, Audit trail

Read: AI evidence packs: a primer for compliance teams, The evidence pack

Back to the list of terms

GuardrailsCommon term

The controls placed around a model or an agent to keep it within bounds: input and output filters, refusal rules in its instructions, limits on what its tools can do, approval steps before an action. They are a design choice, not evidence: a test shows whether they hold under pressure.

See also: Jailbreak, Human oversight, Tool call

Read: Agentic AI under the EU AI Act: control framework

Back to the list of terms

JailbreakCommon term

A prompt written to make a model or an agent ignore its own rules: role play, an invented authority, a request split over several turns. A system that refuses a direct request can still give in to a rephrasing, which is why tests apply several pressure techniques and each test is sent ten times.

See also: Prompt injection, Adversarial testing, Guardrails

Read: How to test an AI agent against the EU AI Act

JudgeVidimus term

The model that grades each of the agent’s replies against the clause, the situation the test set and the expected behaviour. Every verdict quotes the reply it graded and records which model graded it and the version of its instructions, so a person can check it. The judge is not told which model or vendor runs your agent and never sees its system prompt. Every failure on a critical clause gets a second model’s opinion, recorded beside the verdict, and a borderline result goes to your reviewer.

See also: Probe (test), Pass rate, Attempt

Read: How we test

Back to the list of terms

MCP (Model Context Protocol)Common term

An open protocol through which an AI application connects to tools and data sources: a server lists the tools it offers and the agent calls them. Vidimus can test an agent exposed as an MCP server, list a server’s tools during registration, and find agents listed in MCP registries.

See also: A2A (Agent2Agent protocol), Tool call

Read: Integrations

Back to the list of terms

Pass rateVidimus term

The share of an agent’s graded replies to one test that showed the expected behaviour, and the finding Vidimus reports for that test: Passed at 90% or more, Needs review at 70% or more, Failed below that, and Inconclusive when fewer than 8 of the 10 attempts could be graded. It is printed with its counts, such as “8 of 9 (89%)”.

See also: Attempt, Probe (test), Judge

Read: How we test

Probe (test)Vidimus term

One test, sent to your agent ten times by default: the article it puts to the test, the behaviour we expect, and how often the agent showed that behaviour, reduced to one outcome (passed, needs review, failed or inconclusive). An agent can answer the same message differently from one time to the next, which is why one reply is never the finding. “Test” and “probe” mean the same thing on this site, and probe is the product’s own word.

See also: Attempt, Pass rate, Judge, Adversarial testing

Read: How to test an AI agent against the EU AI Act, How we test

Prompt injectionCommon term

An attack that hides instructions in content an agent reads, such as an email, a web page, a document or a tool’s output, so that the agent follows them instead of its own. For an agent that can call tools, it is how a stranger makes it pay, send or delete. Vidimus tests it as one of its pressure techniques, injected content.

See also: Jailbreak, Guardrails, Tool call

Read: How to test an AI agent against the EU AI Act

Back to the list of terms

Red teamingCommon term

A security exercise in which a team attacks a system as an adversary would, to find weaknesses before someone else does. Red-teaming platforms for AI find jailbreaks and data leaks but do not map findings to the articles of a regulation, which is what a regulatory test plan adds.

See also: Adversarial testing

Read: Compare

Review priorityVidimus term

Vidimus’s own ranking of how much review attention an agent needs, from its registration answers. Provisional, and not a legal determination. Its steps are Routine, Elevated, Significant, Severe. It is never printed in the regulatory sections of an evidence pack, and it is not the AI Act classification, which follows the rules of the Act.

See also: Classification record, High-risk AI system

Back to the list of terms

Test pattern libraryVidimus term

The library of test designs the test plan draws from: how to ask, how to push, what counts as a pass. A design is written once and applied to each obligation it fits, so the same pressure is applied the same way to every agent. The evidence pack records which version of the library a test run used.

See also: Corpus, Probe (test)

Test runVidimus term

An agent’s test plan, sent to the live agent: every test, each sent ten times. A run sends at most 150 tests. When the plan holds more, it keeps the most severe and names the ones it left out. A run can be cancelled at any time, and repeated on a schedule, with an alert when the pass rate drops.

See also: Probe (test), Attempt, Test pattern library

Read: How to test an AI agent against the EU AI Act, How we test

Tool callCommon term

An action an agent takes through an interface rather than in words: issuing a payment, writing to a record, sending a message, calling a search. Vidimus reads the calls from the agent’s reply, its tool log or its event trace, so an agent that says it referred a case but never called the referral tool is caught by the recorded tool calls. A call to a tool the agent never declared is a finding.

Read: Agentic AI under the EU AI Act: control framework

Back to the list of terms