EU Regulation | AI Governance

EU AI Act: Cybersecurity Requirements for AI Systems.

The EU AI Act (Regulation EU 2024/1689) is the world's first binding AI law. It classifies AI systems by risk, mandates cybersecurity measures, and imposes fines of up to EUR 35 million. It applies to any company placing AI on the EU market - regardless of where the company is based.

Schedule AI Security Assessment Learn about AI Penetration Testing

Last updated: March 2026 - reviewed by certified experts

Feb. 2025
Prohibited practices in force: 35 M EUR
Maximum fine: Dec. 2027
High-Risk obligations (Annex III): 4 risk classes
From minimal to prohibited

Overview

What is the EU AI Act?

The EU AI Act (Regulation EU 2024/1689) is the world's first binding AI regulation. The regulation was published in the Official Journal of the EU on 13 June 2024 and applies directly in all EU member states - without national implementing legislation.

The AI Act follows a risk-based regulatory model: the higher the potential harm of an AI system, the stricter the requirements. Particularly relevant for cybersecurity are the requirements on robustness and resilience under Art. 15, and the obligation to conduct adversarial testing (red teaming) for GPAI models with systemic risk under Art. 55.

The AI Act has extraterritorial reach: it applies to any provider placing an AI system on the EU market, regardless of where the provider is headquartered. This makes it relevant for companies in the US, UK, Asia and beyond.

EU AI Act at a Glance

EU Regulation: EU 2024/1689, 13 June 2024
Bans from: 2 February 2025
GPAI rules from: 2 August 2025
High-Risk from: 2 December 2027Annex I: 2 August 2028
DE Authority: Bundesnetzagentur + BSI (Cyber)
Max. fine: EUR 35 M / 7% global turnover

High-Risk deadline approaching

Providers of high-risk AI systems must fulfil all conformity obligations by December 2027 (Digital Omnibus Regulation (EU) 2026/1744) - conformity assessment, CE marking, EU database registration.

Risk-Based Approach

The 4 Risk Classes of the AI Act

The AI Act classifies AI systems by their risk potential. The higher the risk class, the stricter the requirements - from no obligations to a complete prohibition. The classification determines which cybersecurity and compliance measures you need to take.

Risk Class

Obligations

Examples

Prohibited

AI systems posing unacceptable risks, fully prohibited as of 02 February 2025.

Social scoring by public authorities or private entities
Emotion recognition in the workplace and educational institutions
Biometric categorisation to infer sensitive characteristics
Manipulation of unconscious behaviour (subliminal techniques)
Real-time remote biometric identification in public spaces (exceptions for law enforcement)
Exploitation of vulnerabilities of specific groups

High Risk

AI systems listed in Annex III with extensive obligations from 02 December 2027; product-integrated Annex I systems from 02 August 2028 (Art. 113 AI Act as amended by Regulation (EU) 2026/1744).

Credit scoring and creditworthiness assessment
HR systems: recruitment and promotion decisions
Biometric identification and categorisation
Critical infrastructure operation: power, water, gas networks
Law enforcement, border control, justice
Education and vocational training
Migration and asylum

Limited Risk

Transparency obligations: users must know they are interacting with an AI system.

Chatbots and conversational AI systems
Deepfakes and AI-generated content (labelling required)
Emotion recognition systems (outside prohibited areas)

Minimal Risk

No specific obligations - voluntary codes of conduct recommended.

Spam filters and email classification
Recommendation systems (without significant risks)
AI in video games
Simple image-processing applications

Which risk class applies to your AI system?

The classification determines your compliance effort. AWARE7 assesses your AI systems definitively and creates an action plan based on your specific risk class.

Request classification

Art. 15 AI Act

Cybersecurity Obligations for AI Systems

Art. 15 AI Act requires providers of high-risk AI systems to ensure robustness, resilience and cybersecurity. The requirements apply throughout the entire lifecycle - from development to decommissioning.

Robustness against Adversarial Attacks

High-risk AI systems must be resilient against hostile manipulation (Art. 15(1) AI Act). This includes robustness tests against adversarial examples, data poisoning and model inversion. Technical measures must be documented throughout the entire lifecycle.

AI Penetration Testing

Data Integrity and Data Quality

Training, validation and test datasets must be appropriate in terms of errors, completeness, representativeness and statistical properties (Art. 10 AI Act). Data governance processes and bias checks are mandatory.

Technical Documentation and Logging

Complete recording of all relevant events throughout the entire operational period (Art. 12 AI Act). Logs must enable traceability and comprehensibility of the AI system - for audits, authorities and law enforcement.

Security Consulting

Human Oversight

High-risk AI systems must be designed so that natural persons can effectively monitor outputs (Art. 14 AI Act). Human-in-the-Loop or Human-on-the-Loop - depending on the risk profile. Override mechanisms must be technically embedded.

Accuracy, Precision and Bias Testing

AI systems must be operated with appropriate accuracy, robustness and cybersecurity throughout their entire lifecycle (Art. 15 AI Act). Regular bias tests and performance monitoring are mandatory, especially for systems making decisions about individuals.

AI Security Testing

Conformity Assessment and Registration

Providers of high-risk AI systems must conduct a conformity assessment before placing on the market (Art. 43 AI Act) and register the system in the EU database (Art. 49 AI Act). CE marking is required.

Incident Reporting to Authorities

Serious incidents and malfunctions of high-risk AI systems must be reported to the competent national market surveillance authority (Art. 73 AI Act). In Germany this will likely be the Federal Network Agency (Bundesnetzagentur) in coordination with the BSI.

Red Teaming for GPAI Models

Providers of general-purpose AI (GPAI) models with systemic risk must conduct adversarial testing (Art. 55 AI Act). Red-teaming tests uncover security vulnerabilities, abuse potential and undesirable behaviour.

Red Teaming for AI

Note: The AI Act's cybersecurity requirements for high-risk Annex III systems apply from December 2027 (Digital Omnibus Regulation (EU) 2026/1744). Providers of GPAI models with systemic risk must be able to demonstrate adversarial red teaming from August 2025 (Art. 55 AI Act). AWARE7 recommends starting preparation immediately.

Art. 50 AI Act

Labelling AI content: the EU icons for AI-generated content

The transparency obligations of Art. 50 AI Act have applied since 2 August 2026. Deployers of generative AI systems must disclose two types of content: deep fakes - AI-generated or AI-manipulated image, audio or video content that resembles existing persons, objects, places or events and would falsely appear authentic or truthful - and AI-generated or AI-manipulated text published to inform the public on matters of public interest - unless it has undergone human review and a natural or legal person holds editorial responsibility. To support this disclosure, the European Commission published a freely usable icon set in 2026 as part of Section 2 of the Code of Practice on marking and labelling of AI-generated content.

Basic "AI" icon

When AI was involved in creating deep fake content or published text and a custom text label or an interactive second layer provides the details.

Example
A deep fake video with the text label "voices generated with" followed by the icon

"AI Generated"

When the content is fully generated by AI - with no human-created elements and no human editorial control apart from prompting.

Example
Fully AI-generated deep fake videos, AI-composed music, AI-generated news summaries

"AI Modified"

When pre-existing, human-made content was partially modified with AI, turning it into a deep fake or text on matters of public interest.

Example
A face in an authentic photograph is swapped with another person's face using AI

Exceptions and reduced obligations

Editorial responsibility: AI-generated text is exempt when both conditions are met: it has undergone human review or editorial control, and a person or organisation holds editorial responsibility for its publication.
Art and satire: not a full exemption - for deep fakes in evidently artistic, creative, satirical or fictional works, the obligation is limited to an appropriate disclosure that does not hamper the display or enjoyment of the work.
No deep fakes: illustrative AI images that do not make existing persons, objects, places or events falsely appear authentic do not fall under Art. 50(4).
Law enforcement: use authorised by law to detect, prevent, investigate or prosecute criminal offences is exempt.

Placement rules for the icons

Art. 50(5) AI Act legally requires clear, distinguishable and accessible information at the latest at the time of first exposure. The following detailed rules come from Section 2 of the Code of Practice and bind its signatories:

Clearly perceivable and distinguishable at the latest at the time of first exposure, with no intervening overlay elements.
Directly embedded into the content (except for creative works), unless equivalent alternatives such as a user interface overlay are available - the label must remain visible when content is reshared or downloaded.
Accessible: clearly visible size, plain language in accompanying labels, alt text or ARIA labels for assistive technologies.
Time-limited disclosures must remain visible long enough to be read and understood.

Important: using the EU icons is optional - the labelling obligation under Art. 50(4) AI Act is not. The icon alone does not establish legal compliance; deployers remain responsible for ensuring that their disclosure meets the legal requirements. Separate from this, providers of generative AI systems must mark their outputs in a machine-readable format (Art. 50(2) AI Act). The icons are available for download as SVG and PNG from the European Commission (as of 20 July 2026).

Regulatory Landscape

AI Act + NIS-2 + GDPR + CRA: The Full Regulatory Picture

For companies deploying AI, a complex interplay of several EU frameworks arises. All must be observed in parallel - there is no choice.

EU AI Act

Risk classes, robustness, red teaming, logging, human oversight. AI-specific regulation.

Applies from 2025/2026

NIS-2

NIS-2 Directive

Cybersecurity of the entire infrastructure including AI systems. Applies to affected sectors across the EU.

NIS-2 Guide

GDPR

Data protection for AI systems processing personal data. Applies to virtually all AI applications.

GDPR Guide

CRA

Cyber Resilience Act (CRA)

Cybersecurity requirements for products with digital elements - includes embedded AI.

CRA Guide

Note for non-EU companies: The AI Act applies to any provider placing AI on the EU market regardless of their location - similar to the GDPR's extraterritorial reach. Importers and distributors in the EU share responsibility for the compliance of the AI systems they market.

Case Studies

Real AI Incidents: What the AI Act Would Have Prevented

These cases demonstrate why the AI Act is necessary - and what consequences uncontrolled AI deployment can bring. Under the AI Act, many of these incidents would have been regulatorily addressed or prevented.

March 2023

ChatGPT & GDPR - Data Leak at OpenAI

OpenAI reported a bug in the open-source Redis library that allowed users to see the chat titles of other users. Italy's data protection authority Garante imposed a temporary ban on ChatGPT and initiated proceedings for GDPR violations. The incident illustrates that AI systems are not closed systems - data leaks can arise from unexpected dependencies.

Ongoing since 2022

Clearview AI - Biometric Data Without Consent

US AI provider Clearview AI was fined by several EU data protection authorities (France: EUR 20 million, Italy: EUR 20 million, Greece: EUR 20 million). Clearview had scraped biometric data from the internet without the consent of those affected. Under the AI Act, this practice would be classified as prohibited biometric categorisation.

October 2018

Amazon HR AI - Discrimination in Hiring

Amazon's internal AI recruitment system systematically favoured male candidates because it was trained on historically male-dominated hiring data. Amazon discontinued the project. Under the AI Act, such an HR system would be classified as high-risk and would have required extensive bias tests and human oversight - before deployment.

Ongoing since 2021

Tesla Autopilot - Regulatory Investigations

US regulator NHTSA investigated over 750 accidents connected to Tesla's Autopilot system. In the EU, autonomous driving systems would be classified as high-risk AI (Annex III, transport). The AI Act requires providers to carry out extensive conformity assessments, log every incident and conduct regular safety reviews.

Advisory Services

How AWARE7 Supports AI Act Compliance

From risk classification to adversarial testing - AWARE7 guides you through all AI Act requirements with a focus on technical cybersecurity.

AI Penetration Testing & Red Teaming

Adversarial testing of your AI systems: adversarial examples, data poisoning, model extraction, prompt injection - in line with Art. 15 and Art. 55 AI Act.

Request AI pentest

AI Act Gap Analysis & Classification

Definitive risk classification of your AI systems, gap analysis against all AI Act requirements, prioritised action plan with timeline.

Start gap analysis

Integrated Compliance: AI Act + NIS-2 + GDPR

Holistic advice combining AI Act, NIS-2 and GDPR in a single engagement - rather than three parallel projects.

Understand NIS-2 interaction

„The EU AI Act is not abstract regulation - it has concrete cybersecurity requirements that demand technical expertise. Anyone deploying AI in safety-critical domains needs adversarial testing, not just legal advice.“

Chris Wojzechowski

Auditor with additional audit procedure competence (Section 8a BSIG) · AWARE7 GmbH

FAQ

Frequently Asked Questions about the EU AI Act

The most important questions about the EU AI Act - answered with technical depth and practical focus.

When does the EU AI Act apply?

The EU AI Act (Regulation EU 2024/1689) entered into force on 1 August 2024 and has been generally applicable since 2 August 2026. It applies in stages: prohibited AI practices have been in effect since 2 February 2025, rules for general-purpose AI models (GPAI) since 2 August 2025 and the Art. 50 transparency obligations since 2 August 2026. The Digital Omnibus Regulation (EU) 2026/1744 (in force since 27 July 2026) postponed the high-risk obligations: Annex III systems are covered from 2 December 2027, product-integrated Annex I systems from 2 August 2028. The AI Act applies directly in all EU member states and has extraterritorial reach: non-EU companies placing AI systems on the EU market must comply.

Which AI systems are classified as "high-risk" under the AI Act?

High-risk AI systems are listed exhaustively in Annex III of the AI Act. These include: biometric identification and categorisation, operation of critical infrastructure (energy, water, transport), education and vocational training, employment and human resources management (HR systems), access to private and public services (credit scoring, social benefits), law enforcement and border control, migration and asylum proceedings, and the administration of justice. Additionally: AI systems in safety-relevant products subject to EU harmonisation legislation (Annex I, e.g. medical devices).

What does the AI Act mean for companies using AI tools such as ChatGPT?

Companies that only use AI services (deployers/operators) have fewer obligations than providers (developers). Nevertheless they must: ensure that AI systems deployed are compliant (due diligence), inform employees when AI decisions affect them, ensure human oversight for high-risk applications, and for some systems carry out a data protection impact assessment under GDPR. Important: anyone who fine-tunes or adapts a GPAI model for their own purposes may become a provider themselves.

How high are the fines under the EU AI Act?

The AI Act provides for three fine categories: violations of prohibited AI practices (Art. 5): up to EUR 35 million or 7% of global annual turnover. Violations of high-risk obligations and GPAI rules: up to EUR 15 million or 3% of global annual turnover. Inaccurate information to authorities: up to EUR 7.5 million or 1.5% of annual turnover. In each case the higher amount applies. Proportionally lower caps apply to SMEs and start-ups.

What is a GPAI model and what obligations apply?

GPAI (General Purpose AI) are AI models trained for a wide variety of tasks and usable across different application contexts - typically large language models such as GPT-4, Gemini or LLaMA. From 2 August 2025, all GPAI providers must provide: technical documentation, a summary of training data (copyright), and EU code-of-practice compliance. For GPAI models with systemic risk (>10^25 FLOPs training) additional requirements apply: adversarial red-teaming (Art. 55), cybersecurity obligations, reporting of serious incidents to the EU AI Office.

How do the AI Act, GDPR and NIS-2 interact?

The three frameworks complement and overlap each other: the GDPR governs data protection for AI systems that process personal data - which applies to almost all practical AI applications. NIS-2 requires affected entities to secure their entire IT infrastructure, including AI systems. The AI Act adds AI-specific security requirements (robustness, adversarial attacks, human oversight). Companies must comply with all three simultaneously. AWARE7 provides integrated advice: AI Act compliance plus NIS-2 plus GDPR from a single source.

What is BSI AIC4 and how does it relate to the AI Act?

The German Federal Office for Information Security (BSI) has developed the AI Cloud Service Compliance Criteria Catalogue (AIC4), a catalogue of criteria for the security of AI cloud services. AIC4 addresses security aspects such as data integrity, model robustness, transparency and explainability - covering significant parts of the AI Act requirements under Art. 15 (robustness) and Art. 12 (logging). An AIC4 audit by an accredited testing laboratory can serve as evidence of AI Act compliance where requirements overlap. AWARE7 supports AIC4 audits as preparation for AI Act compliance.

Do I need to register my AI systems in an EU database?

Yes - providers of high-risk AI systems under Annex III must register their system in the EU-operated database before placing it on the market or putting it into service (Art. 49 AI Act). Deployers of high-risk AI systems in certain areas (e.g. public administration) must also be registered in the EU database. Registration serves market transparency and facilitates regulatory supervision. Exceptions apply to law enforcement authorities in certain constellations.

How does an AI security assessment at AWARE7 work?

Our AI security assessment follows four phases: (1) Classification - determining the risk class of your AI systems under Art. 5, Art. 6 and Annex III of the AI Act. (2) Gap analysis - comparison against all relevant AI Act requirements (data governance, robustness, logging, human oversight, documentation). (3) Technical testing - adversarial testing (adversarial examples, data poisoning, model extraction) and bias analysis. (4) Report and action plan - documentation for conformity assessment, CE marking and authorities. Typical duration: 4 to 10 weeks.

Does the AI Act also apply to AI systems developed outside the EU?

Yes - the AI Act has extraterritorial effect, similar to the GDPR. It applies to: providers placing AI systems on the EU market (regardless of the provider's location), deployers within the EU, and AI systems whose outputs are used in the EU. This means: US, Chinese or Israeli AI providers marketing their products in the EU must comply with the AI Act. Importers and distributors in the EU share responsibility for the compliance of the systems they market.

Which AI-generated content must be labelled under Art. 50 AI Act?

The disclosure obligation for deployers (Art. 50(4) AI Act) covers two cases: (1) Deep fakes - AI-generated or AI-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. (2) AI-generated or AI-manipulated text published with the purpose of informing the public on matters of public interest. Important exceptions: for text, the obligation does not apply where the content has undergone human review or editorial control and a natural or legal person holds editorial responsibility for its publication. For evidently artistic, creative, satirical or fictional works, disclosure is limited to a manner that does not hamper the display or enjoyment of the work. Illustrative AI images without deep fake character do not fall under Art. 50(4). Separate from this, providers must mark outputs of generative AI systems in a machine-readable format (Art. 50(2)). The transparency obligations have applied since 2 August 2026.

What are the EU icons for labelling AI-generated content?

Since 2026, the European Commission provides three icons that deployers of generative AI systems may use to label content subject to disclosure: a basic "AI" icon, an "AI Generated" icon for fully AI-generated content and an "AI Modified" icon for partially AI-modified content. The icons are part of Section 2 of the Code of Practice on marking and labelling of AI-generated content, have been user-tested and are freely available in four variations (black, white, each also with 50 percent transparency) as SVG and PNG - no attribution required. Important: using the icons is optional, the labelling obligation under Art. 50(4) AI Act is not. The icon alone does not establish legal compliance - deployers remain responsible for ensuring that their disclosure meets the requirements of Art. 50.

Aus dem Blog

Weiterführende Artikel

Alle Artikel

Threat Intelligence 04. März 2026

KI-Cyberangriffe: Deepfakes, AI-Phishing und automatisierte Exploits

KI-gestützte Angriffe: Deepfake-Audio für BEC, LLM-Spear-Phishing und AI-Malware - Erkennungsmethoden und Integration in die Security-Strategie.

Jan Hörnemann

10 Min.

Threat Intelligence 04. März 2026

LLM Security: Prompt Injection, Jailbreaking und KI-Red-Teaming

LLM Red Teaming: Prompt Injection, Jailbreaking, Training-Data-Poisoning und OWASP Top 10 for LLM Applications - mit Defense-Strategien.

Jan Hörnemann

12 Min.

Offensive Security 12. August 2019

GDPiRate: Seitenkanalangriff deckt DSGVO-Sicherheitslücken auf!

Seit Mai 2018 haben Nutzer das Recht online auf persönliche Daten zuzugreifen. Drei Studien decken nun DSGVO-Sicherheitslücken auf.

Matteo Große-Kampmann

4 Min.

Alle Artikel

Schedule an AI Security Assessment

In a free 30-minute call, we analyse your AI systems, determine the relevant risk class under the AI Act, and show which cybersecurity measures are specifically required - with a timeline and fixed-price proposal.

Book AI Assessment Now Learn about AI Penetration Testing

Free · 30 minutes · No obligation