EU AI Act

The EU AI Act (Regulation (EU) 2024/1689) entered into force on 1 August 2024 and classifies AI systems by risk: prohibited practices, high-risk AI systems, systems subject to transparency duties, and general-purpose AI models. From an information security perspective, the high-risk category matters most – AI used as a safety component of regulated products (Annex I) and the use cases listed in Annex III, such as critical infrastructure, employment and worker management, creditworthiness assessment or biometric identification. The obligations apply in stages: prohibitions and AI literacy since February 2025, general-purpose AI model duties since August 2025, and the high-risk requirements from August 2026, or August 2027 for Annex I products. That timeline is politically contested, however: under its Digital Omnibus package the European Commission has proposed postponing some high-risk deadlines, and the legislative process is not yet complete. Organisations should therefore verify the current legal position before fixing any deadline plan rather than counting on a delay.

Article 15 is the core security provision. High-risk AI systems must be designed and developed so that they achieve an appropriate level of accuracy, robustness and cybersecurity, and maintain those properties across their entire lifecycle. Accuracy metrics have to be declared in the instructions for use; the system must be resilient against errors, faults and inconsistencies – for example through redundancy or fail-safe designs – and where a system continues to learn after deployment, feedback loops must not produce biased outputs. Cybersecurity here explicitly goes beyond conventional IT hardening: the regulation names AI-specific attack vectors such as data poisoning and model poisoning, adversarial examples or model evasion, confidentiality attacks on training data and models, and the exploitation of model flaws. Article 15 is supported by the risk management system under Article 9, data governance under Article 10, automatic logging under Article 12, human oversight under Article 14 and the provider's quality management system under Article 17. Article 42(2) adds a presumption of conformity: high-risk AI systems certified under a European cybersecurity certification scheme pursuant to Regulation (EU) 2019/881 are presumed to comply with the cybersecurity requirements of Article 15 to that extent.

In practice this means the AI Act does not call for a second, separate security management system – it can largely be embedded into an existing ISMS based on ISO/IEC 27001. Risk assessment, protection requirement analysis, supplier and service provider management, change and patch management, logging and monitoring, vulnerability management and incident response already provide the framework. What has to be added are AI-specific assets (training and validation data, models, weights, prompts, vector databases, MLOps pipelines), AI-specific threat scenarios and an AI inventory that assigns every system a role – provider or deployer – and a risk class. ISO/IEC 42001 offers a dedicated AI management system that is structurally compatible with ISO/IEC 27001, while harmonised European standards are being developed by CEN/CENELEC JTC 21. Where an organisation is already in scope of NIS2 or the Cyber Resilience Act, the evidence should be maintained jointly to avoid duplicated effort. Breaches of the high-risk requirements can be fined up to EUR 15 million or 3 per cent of worldwide annual turnover under Article 99. In Germany, the national supervisory and market surveillance structures had not yet been finally settled in law, with a coordinating role foreseen for the Federal Network Agency (Bundesnetzagentur).