In depth

AI Act: a guide to the EU artificial intelligence regulation and cybersecurity

What Regulation (EU) 2024/1689 on artificial intelligence requires: dates as updated by the Digital Omnibus, risk levels, obligations for high-risk systems and general-purpose models and, above all, the cybersecurity requirements, AI-specific attacks, serious incidents and the relationship with the CRA, NIS 2 and technical standards, with references to the articles.

Checked against the official text on 16 November 2026 · How we work

In short for businesses

Who it concerns
You develop or sell artificial intelligence systems, or you use them in your business: in that case you are a “deployer”. Using a chatbot at work counts as use too.
What to do
For most companies: do not use the prohibited practices, make sure the people who use AI have the right skills, and meet the transparency obligations, meaning people must know when they are talking to a chatbot or seeing AI-generated content (Art. 50). Much heavier obligations, including cybersecurity, apply to high-risk systems, for example those used to recruit staff or in critical infrastructure.
When
Prohibitions from 2 February 2025, transparency from 2 August 2026, high-risk systems from 2 December 2027 (Annex III) and from 2 August 2028 (Annex I).

A simplified summary: details and exceptions are below. Terms with a dotted underline link to the glossary.

Regulation (EU) 2024/1689 on artificial intelligence, known as the AI Act, is the world’s first comprehensive law on AI. It lays down harmonised rules for placing on the market, putting into service and using AI systems in the Union, with a risk-based approach: some practices are banned, high-risk systems must meet precise requirements, other systems have transparency obligations, and general-purpose AI models have their own rules (Article 1). It entered into force on 1 August 2024 and applies in stages until 2028.

It is a product regulation, built like the other EU product safety laws: requirements laid down in law, harmonised standards that give a presumption of conformity, notified bodies, an EU declaration of conformity and the CE marking. Regulation (EU) 2026/1744 (the Digital Omnibus on AI), in force since 27 July 2026, postponed the obligations for high-risk systems and changed several other rules.

The AI Act is not a cybersecurity law, but cybersecurity is one of its pillars: an AI system that an attacker can manipulate is not trustworthy, and a very capable model can itself become an attack tool.

Accuracy, robustness and cybersecurity (Article 15). High-risk AI systems must be designed and developed to achieve an appropriate level of accuracy, robustness and cybersecurity and to perform consistently in those respects throughout their lifecycle (paragraph 1). In particular, they must:

  • be as resilient as possible to errors, faults and inconsistencies, including through redundancy, backup or fail-safe plans (paragraph 4);
  • if they continue to learn after being placed on the market, limit feedback loops, where biased outputs become inputs for future operations (paragraph 4);
  • be resilient against attempts by unauthorised third parties to alter their use, outputs or performance by exploiting system vulnerabilities (paragraph 5);
  • use technical cybersecurity solutions that are appropriate to the relevant circumstances and risks (paragraph 5).

AI-specific vulnerabilities. The regulation requires measures to prevent, detect, respond to, resolve and control the attacks typical of AI listed in the table below (Article 15(5)). Attacks can target AI-specific assets such as training data and models, but also the system’s digital assets or the underlying ICT infrastructure: providers must therefore put in place security controls that also take the infrastructure into account (recital 76).

Other requirements that serve security. Article 15 relies on the other requirements for high-risk systems: the risk management system throughout the lifecycle (Article 9), governance of training, validation and testing data (Article 10), automatic recording of events (logs), which makes it possible to reconstruct what happened (Article 12), and human oversight, which must allow the system to be stopped safely (Article 14).

Models with systemic risk. Providers of the most capable general-purpose AI models must ensure an adequate level of cybersecurity protection for the model and its physical infrastructure (Article 55(1)(d)), carry out adversarial testing and report serious incidents to the AI Office.

Serious incidents. Providers of high-risk systems must report serious incidents to the authorities, including the serious and irreversible disruption of critical infrastructure, with deadlines as short as two days (Article 73).

Presumption of conformity through the CRA and the Cybersecurity Act. High-risk systems that are also products with digital elements and comply with the Cyber Resilience Act are deemed to meet the cybersecurity requirements of Article 15 (Article 42(3) and Article 12 CRA); the same applies to systems certified under a European cybersecurity certification scheme (Article 42(2)).

Back to contents

AI-specific attacks

Attack What it does Where it is mentioned
Data poisoning Manipulates the training data so the model learns wrong or hidden behaviour Article 15(5); recitals 76 and 77
Model poisoning Tampers with pre-trained components used in training Article 15(5)
Adversarial examples, model evasion Crafts inputs designed to make the model make a mistake Article 15(5); recitals 76 and 77
Confidentiality attacks, membership inference Extracts information about the training data from the model Article 15(5); recital 76
Model flaws Exploits errors or weaknesses of the model itself Article 15(5)
Model leakage, theft or unauthorised release, circumvention of safety measures Targets the weights, algorithms, servers and datasets of models with systemic risk Article 55(1)(d); recital 115

For models with systemic risk, protection can include securing weights, algorithms, servers and datasets, operational security measures, cybersecurity policies and cyber and physical access controls (recital 115).

Back to contents

Key dates

Date What happens
12 July 2024 Publication in the Official Journal of the EU
1 August 2024 Entry into force (Article 113)
2 February 2025 General provisions, AI literacy and prohibited practices (Chapters I and II)
2 August 2025 Notified bodies, general-purpose AI models, governance, penalties and confidentiality (Article 113(b))
27 July 2026 The Digital Omnibus on AI, Regulation (EU) 2026/1744, enters into force (Article 4)
2 August 2026 Most of the regulation applies: transparency obligations (Article 50), Commission fines on general-purpose model providers (Article 101), market surveillance
2 December 2026 New bans on non-consensual intimate images and child sexual abuse material; deadline for marking the content of generative systems already on the market (Articles 5, 111 and 113, as amended)
2 August 2027 Compliance of general-purpose models placed on the market before 2 August 2025 (Article 111(3)); national AI regulatory sandboxes operational (Article 57)
2 December 2027 Obligations for Annex III high-risk systems, including critical infrastructure (Article 113(c)(i))
2 August 2028 Obligations for Annex I high-risk systems, i.e. safety components of regulated products (Article 113(c)(ii))
2 August 2030 Deadline for high-risk systems already on the market that are used by public authorities (Article 111(2))
31 December 2030 Deadline for AI systems in the large-scale IT systems listed in Annex X (Article 111(1))

Before the omnibus, the obligations for high-risk systems were due to apply on 2 August 2026 (Annex III) and 2 August 2027 (Annex I). High-risk systems already on the market before the date of application stay outside the regulation until their design undergoes significant changes (Article 111(2)). The dates are also in the document record.

Back to contents

Who it applies to and the risk levels

In practice: the first step is to work out which level your system falls into: most obligations concern high-risk systems, including AI used as a safety component of critical infrastructure.

The regulation applies to providers placing AI systems or models on the EU market, wherever they are established, to deployers (those using an AI system under their authority) established in the Union, to importers and distributors, and to providers and deployers in third countries when the system’s output is used in the Union (Article 2(1)). Military, defence and national security uses, pure scientific research and purely personal non-professional use are excluded (Article 2).

Level What it covers Rules
Prohibited practices Manipulation and exploitation of vulnerabilities, social scoring, some forms of biometric identification and, from 2 December 2026, generating non-consensual intimate images and child sexual abuse material Ban (Article 5)
High risk Safety components of Annex I products subject to third-party assessment (Article 6(1)) and Annex III use cases, including biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration and justice (Article 6(2)) Requirements and conformity assessment (Chapter III)
Transparency risk Chatbots, synthetic content, deepfakes, emotion recognition Inform people and mark content (Article 50)
General-purpose models Models trained on large amounts of data and able to perform many tasks Documentation, copyright policy, training data summary (Article 53); further obligations if they have systemic risk (Article 55)

For cybersecurity, point 2 of Annex III matters most: AI systems used as safety components in the management and operation of critical digital infrastructure, road traffic and the supply of water, gas, heating and electricity are high-risk. An Annex III system is not high-risk if it does not pose a significant risk, for example because it only performs a narrow procedural task (Article 6(3)). Under the omnibus, AI systems used only for non-safety purposes such as user assistance, performance optimisation or quality control are not safety components, unless their failure would endanger health and safety (Article 6(1a) and (1b)).

Back to contents

Obligations for high-risk systems

In practice: the provider of a high-risk system documents it, assesses conformity, affixes the CE marking and registers it; the deployer follows the instructions, oversees it and keeps the logs for at least six months.

The provider of a high-risk system (Articles 16-22):

  1. complies with the requirements of Articles 8-15, including Article 15 on accuracy, robustness and cybersecurity;
  2. has a quality management system (Article 17), which also covers data management, risk management and post-market monitoring;
  3. draws up the technical documentation (Annex IV), which also describes the cybersecurity measures; SMEs and small mid-caps can use a simplified form (Article 11);
  4. keeps the automatically generated logs when they are under its control (Article 19);
  5. carries out the conformity assessment, draws up the EU declaration of conformity, affixes the CE marking and registers the system in the EU database (Articles 43, 47, 48 and 49);
  6. after sale, keeps a post-market monitoring plan (Article 72), corrects or withdraws non-compliant systems and reports serious incidents (Article 73).

The deployer uses the system according to the instructions, assigns human oversight to competent people, monitors its operation, keeps the logs under its control for at least six months and informs the provider and the authorities if the system presents a risk or causes a serious incident (Article 26). Some deployers, such as public bodies, must assess the impact on fundamental rights (Article 27).

Market surveillance authorities have full access to the documentation and datasets, including remotely through APIs, and can request the source code when other checks are not enough (Article 74(12) and (13)).

Back to contents

General-purpose models with systemic risk

A general-purpose AI model has systemic risk if it has high-impact capabilities: this is presumed when the compute used for training exceeds 10^25 floating-point operations; the Commission can also designate it by decision (Article 51). On top of the obligations for all models (Article 53), providers of these models (Article 55(1)):

  • evaluate the model with state-of-the-art protocols and tools, including documented adversarial testing, to identify and mitigate systemic risks;
  • assess and mitigate systemic risks at Union level;
  • keep track of serious incidents and report them without undue delay to the AI Office, with corrective measures;
  • ensure an adequate level of cybersecurity protection for the model and its physical infrastructure.

Until harmonised standards exist, providers can show compliance by adhering to a code of practice (Article 55(2)). The General-Purpose AI Code of Practice (opens in a new tab), published on 10 July 2025, has a safety and security chapter for models with systemic risk; the Commission and the AI Board have confirmed it is an adequate tool. Supervision of the models lies with the Commission through the AI Office, which can impose fines from 2 August 2026 (Article 101).

Back to contents

Serious incidents

In practice: a serious incident involving a high-risk system is reported to the market surveillance authority within 15 days, or within 2 days if it hits critical infrastructure; an attack may also require CRA and NIS 2 reports.

A serious incident is an incident or malfunction of an AI system that directly or indirectly leads to: the death of a person or serious harm to a person’s health; a serious and irreversible disruption of the management or operation of critical infrastructure; the infringement of Union law obligations protecting fundamental rights; serious harm to property or the environment (Article 3(49)).

Case Reporting deadline (Article 73)
General rule Immediately after establishing a causal link, or its reasonable likelihood, and in any event within 15 days of the provider or deployer becoming aware of it (paragraph 2)
Widespread infringement or serious and irreversible disruption of critical infrastructure Immediately and in any event within 2 days (paragraph 3)
Death of a person Immediately and in any event within 10 days (paragraph 4)

The report goes to the market surveillance authority of the Member State where the incident occurred; an incomplete initial report can be sent first (paragraph 5). The provider investigates without altering the system in a way that could affect the analysis of the causes (paragraph 6). Where the provider is already subject to equivalent reporting obligations, for Annex III systems it reports under the AI Act only infringements of fundamental rights (paragraph 9). Under the omnibus, providers of systems supervised by the AI Office, such as those built on a general-purpose model from the same provider, report to the AI Office (Article 75(1a)). In September 2025 the Commission put out for consultation draft guidance and a reporting template (opens in a new tab).

A cyberattack on an AI system may have to be reported under several laws, to different recipients and on different timelines:

Law What is reported Deadlines To whom
AI Act (Article 73) Serious incidents caused by high-risk systems 2, 10 or 15 days Market surveillance authority
CRA (Article 14) Actively exploited vulnerabilities and severe incidents affecting the security of products with digital elements, from 11 September 2026 24 hours, 72 hours, then a final report Coordinating CSIRT and ENISA, via the single reporting platform
NIS 2 (Article 23) Significant incidents affecting the services of essential and important entities 24 hours, 72 hours, 1 month CSIRT or national authority

Back to contents

AI Act and Cyber Resilience Act

Many AI systems are also products with digital elements covered by the Cyber Resilience Act, which largely applies from 11 December 2027, a few days after the obligations for Annex III systems under the AI Act. The two regulations “complement each other” for the safety and cybersecurity of products (recital 19 of Regulation 2026/1744).

Presumption of conformity. A product with digital elements classified as a high-risk AI system is deemed to meet the cybersecurity requirements of Article 15 if it meets the CRA essential requirements on the product (Annex I, Part I) and on the manufacturer’s processes (Annex I, Part II) and the EU declaration of conformity demonstrates this (Article 12(1) CRA). The omnibus copied the same rule into the AI Act, in Article 42(3), to make it more visible. The requirements on accuracy and robustness, which the CRA does not cover, remain separate.

Conformity assessment. As a rule the AI Act procedure applies (Article 43), and bodies notified under the AI Act can also assess the CRA requirements (Article 12(2) CRA). However, important and critical CRA products that could use internal control alone under the AI Act must follow the CRA procedures for the cybersecurity part, so as not to lower the level of assurance (Article 12(3) CRA and recital 78).

Risk assessment. In the cybersecurity risk assessment carried out for the CRA, the manufacturer must also consider AI-specific risks such as data poisoning and adversarial attacks (recital 77).

Aspect AI Act (EU) 2024/1689 CRA (EU) 2024/2847
What it protects Health, safety and fundamental rights The cybersecurity of the product and its data
Cyber requirements Accuracy, robustness and cybersecurity, AI-specific attacks (Article 15) Essential requirements on the product and on vulnerability handling (Annex I)
After sale Post-market monitoring and reporting of serious incidents (Articles 72 and 73) Vulnerability handling throughout the support period and reporting to ENISA (Articles 13 and 14)
Applies from 2 December 2027 (Annex III) and 2 August 2028 (Annex I) 11 December 2027 (reporting from 11 September 2026)

Back to contents

AI Act, NIS 2, DORA and certification

NIS 2. NIS 2 does not regulate AI systems but organisations: a hospital, an electricity grid operator or a cloud provider using AI remains subject to its cyber risk management and incident notification obligations, including supply chain security. AI systems used as safety components of critical infrastructure, including digital infrastructure, are high-risk under the AI Act (Annex III, point 2), and those using them may have obligations under both laws. See also the in-depth guide to the CER Directive on the resilience of critical entities.

DORA. For banks, insurers and other financial entities, ICT risk management, including AI risks, follows the DORA Regulation; when these entities use high-risk systems, the market surveillance authority under the AI Act is as a rule the financial supervisory authority (Article 74(6)). In June 2026 the European Systemic Risk Board warned that frontier AI models able to discover vulnerabilities and build exploits quickly increase systemic cyber risk in the financial sector (Warning ESRB/2026/3).

Certification. High-risk systems certified, or covered by a statement of conformity, under a European cybersecurity certification scheme of the Cybersecurity Act published in the Official Journal are presumed to meet the cybersecurity requirements of Article 15, to the extent the certificate covers them (Article 42(2)). The first scheme adopted is EUCC for ICT products.

Machinery Regulation. Under the omnibus, the requirements for machinery with high-risk AI will move into the Machinery Regulation through delegated acts, by 2 August 2028: see the in-depth guide to the Machinery Regulation.

Back to contents

The Digital Omnibus on AI

Regulation (EU) 2026/1744 of 8 July 2026, published on 24 July and in force since 27 July 2026, amended the AI Act to simplify its implementation, after delays in preparing technical standards and national authorities (recital 2). The main changes:

  • postponement of the obligations for high-risk systems to 2 December 2027 (Annex III) and 2 August 2028 (Annex I) (Article 113);
  • new bans, from 2 December 2026, on systems generating non-consensual intimate images of identifiable people or child sexual abuse material (Article 5(1)(ba) and (bb));
  • the presumption of conformity through the CRA is also written into the AI Act (Article 42(3));
  • the Machinery Regulation moves from Section A to Section B of Annex I: the AI Act requirements for machinery will be added to the Machinery Regulation through delegated acts;
  • the AI Office has exclusive competence over systems built on general-purpose models from the same provider and over those integrated into very large online platforms and search engines, with inspection and sanctioning powers (Articles 75 to 75d);
  • AI literacy becomes an obligation to take measures, without a guaranteed level, which the Commission and Member States must support (Article 4);
  • simplifications for SMEs and small mid-caps (technical documentation, quality management system, lower fines), exceptional processing of special categories of personal data to correct bias (Article 4a), lighter registration in the EU database for Annex III systems considered not high-risk (Annex VIII) and a single procedure for notified bodies already designated for Annex I products.

Back to contents

Harmonised standards

Those who apply harmonised standards published in the Official Journal are presumed to meet the requirements they cover (Article 40). The Commission requested them from CEN and CENELEC with Implementing Decision C(2025) 3871 of 23 June 2025, which replaced the 2023 request; the work is carried out by the joint technical committee CEN-CENELEC JTC 21.

  • EN 18286:2026, quality management system for the AI Act: the first harmonised standard cited in the Official Journal, by Implementing Decision (EU) 2026/2274 of 12 October 2026;
  • prEN 18282 Artificial Intelligence – Cybersecurity specifications for AI Systems, being drafted in JTC 21 specifically for the cybersecurity requirements of Article 15;
  • ISO/IEC 42001, a certifiable AI management system with the same structure as ISO/IEC 27001;
  • ISO/IEC 23894, guidance on AI risk management;
  • ISO/IEC 27090, published on 14 October 2026: guidance on recognising and mitigating security threats specific to AI systems throughout their lifecycle; it is not a harmonised standard, but it is a useful reference for the measures of Article 15;
  • ETSI EN 304 223, baseline cybersecurity requirements for AI models and systems;
  • the NIST AI RMF, a voluntary US framework for AI risk management.

Where standards are missing or insufficient, the Commission can adopt common specifications through implementing acts (Article 41).

Back to contents

Authorities and penalties

Each Member State designates at least one notifying authority and one market surveillance authority (Article 70). In Italy, Law 132/2025 designated the National Cybersecurity Agency (ACN) as market surveillance authority and single point of contact with the EU, and AgID as notifying authority: supervision of AI is therefore entrusted to the national cybersecurity authority. At EU level the Commission’s AI Office and the Member States’ AI Board operate (Articles 64-65).

Infringement Maximum fine
Prohibited practices (Article 5) EUR 35 million or 7% of worldwide annual turnover, whichever is higher (Article 99(3))
Obligations of providers, deployers, importers, distributors, notified bodies and transparency EUR 15 million or 3% (Article 99(4))
Incorrect, incomplete or misleading information to authorities EUR 7.5 million or 1% (Article 99(5))
General-purpose model providers, fines imposed by the Commission EUR 15 million or 3% (Article 101)

For SMEs, start-ups and small mid-caps the lower of the two amounts applies (Article 99(6) and (6a)). For the systems it supervises, the AI Office can also impose periodic penalty payments of up to 5% of average daily turnover (Article 75c).

Back to contents

Subsequent acts and guidance

Act What it adds
Directive (EU) 2024/2853 on liability for defective products (23 October 2024) Treats software, including AI systems, as a product (recital 13): the manufacturer is liable, without fault, for damage caused by a defective system, assessed taking into account its ability to continue to learn after placing on the market and safety-relevant cybersecurity requirements (Art. 7(2)(c) and (f)). Applies to products placed on the market after 8 December 2026
Guidelines on prohibited AI practices (opens in a new tab) (4 February 2025) How to interpret the bans in Article 5, with examples
Guidelines on the definition of an AI system (opens in a new tab) (6 February 2025) When software is an AI system (Article 3(1))
Implementing Regulation (EU) 2025/454 (opens in a new tab) (7 March 2025) Establishes the scientific panel of independent experts
General-Purpose AI Code of Practice (opens in a new tab) (10 July 2025) Transparency, copyright and a safety and security chapter for models with systemic risk
Guidelines on the obligations of general-purpose AI model providers (opens in a new tab) (18 July 2025) Who is a provider, when a model has systemic risk, open-source exemptions
Draft guidance on serious incidents (opens in a new tab) (26 September 2025) Definitions, examples, reporting template and interplay with other reporting obligations (Article 73(7))
Warning ESRB/2026/3 (opens in a new tab) (25 June 2026) Systemic cyber risks from frontier AI models for the financial sector
Action Plan on Cybersecurity and Artificial Intelligence (7 July 2026) Secure frontier AI usable for defence, AI-ready vulnerability management, fast implementation of NIS 2, DORA and the CRA
Regulation (EU) 2026/1744 (opens in a new tab) (8 July 2026) Digital Omnibus on AI: postponed dates, new bans, CRA presumption in the AI Act, AI Office powers
Implementing Regulation (EU) 2026/1755 (opens in a new tab) (20 July 2026) Evaluations of general-purpose models and Commission sanction proceedings (Articles 92 and 101)
Implementing Decision (EU) 2026/2274 (opens in a new tab) (12 October 2026) First harmonised standard: EN 18286:2026 on the quality management system
Implementing Regulation (EU) 2026/2487 (opens in a new tab) (10 November 2026) Common rules for AI regulatory sandboxes (Article 58): application and selection, sandbox plan, written proof and exit report, extension and suspension; in force from 1 December 2026
Guidelines on the Article 50 transparency obligations (opens in a new tab) (C/2026/6005, 12 November 2026) Who must inform people and mark AI-generated content, deepfakes and texts on matters of public interest, effects of signing up to a code of practice; generative systems placed on the market before 2 August 2026 must apply marking by 2 December 2026

Back to contents

Official sources

Back to contents

This page summarises the text of the act for information only and is not legal advice. Only the text published in the Official Journal of the European Union is authentic.