Up to date as of October 18, 2026. This article takes into account the “Digital Omnibus” regulation on AI published on July 24, 2026. It will be updated as the Commission issues its guidelines.
The European Union has loosened the AI Act timeline for high-risk AI systems. Good news for organizations that were not ready, provided they do not mistake a postponement for a pause: part of the regulation already applies, and the security requirements for AI systems take time to put in place.
A regulation already in force, applied in stages
Regulation (EU) 2024/1689, known as the AI Act, entered into force on August 1, 2024. Its obligations apply gradually:
- since February 2, 2025: prohibition of AI practices deemed unacceptable (social scoring, manipulation, certain forms of biometric identification…) and the first rules on AI literacy for staff;
- since August 2, 2025: obligations for providers of general-purpose AI models, such as large language models;
- since August 2, 2026: the transparency obligations of Article 50, for example informing people that they are interacting with an AI, or disclosing content generated or manipulated by an AI.
What the Digital Omnibus changes
Regulation (EU) 2026/1744 of July 8, 2026, known as the “Digital Omnibus on AI”, was published in the Official Journal of the European Union on July 24, 2026. It amends the AI Act on several points:
- high-risk systems listed in Annex III (recruitment, credit scoring, education, access to certain essential services…): their obligations will apply from December 2, 2027, instead of August 2, 2026;
- high-risk systems embedded in regulated products (Annex I: medical devices, machinery…): from August 2, 2028, instead of August 2, 2027;
- marking of generated content: generative AI systems placed on the market before August 2, 2026 have until December 2, 2026 to comply with the marking obligation;
- AI literacy (Article 4): the obligation is reworded. Providers and deployers must take measures to develop their staff’s skills, without having to guarantee a given level for each person;
- lighter requirements for SMEs and small mid-cap companies, particularly regarding technical documentation.
The transparency obligations of Article 50, however, have not been postponed.
Postponement does not mean pause: the cybersecurity requirement
Article 15 of the AI Act requires high-risk AI systems to achieve an appropriate level of accuracy, robustness and cybersecurity throughout their lifecycle. They must be resilient against attempts by unauthorized third parties to exploit their vulnerabilities in order to alter their use, outputs or performance.
The text explicitly names threats specific to AI:
- poisoning of the training data or of pre-trained components;
- adversarial examples, designed to fool the model;
- attacks on the confidentiality of the data and of the model;
- the exploitation of model flaws.
These protections cannot be improvised in a few weeks. They rest on foundations that security teams know well: asset inventory, risk analysis, control of access to data and models, logging, monitoring and incident response. Organizations subject to NIS2 have every reason to address both topics together: an AI system used in an essential service is also part of the information system to be protected.
What about AI in the SOC?
Security teams increasingly use AI: summarizing and contextualizing alerts, helping with triage, assisting the response. These uses are not, as such, among the high-risk areas of Annex III: the regulation even states that components intended solely for cybersecurity are not safety components of critical infrastructure (recital 55). They still need to be governed:
- confidentiality: security logs contain personal and sensitive data. Where are they sent, and who has access to them? A model hosted in a controlled environment may be preferable;
- prompt injection: a model that reads logs, emails or web pages can receive malicious instructions hidden in that content. It must be treated as untrusted data;
- human oversight: the AI suggests, the analyst decides. Automated actions must stay within validated playbooks;
- traceability: keeping the prompts, the answers and the decisions taken makes it possible to audit the tool and measure its reliability.
Where to start
- Inventory your AI systems and uses, including those built into your SaaS tools and those adopted by teams without approval.
- Classify each use (prohibited practice, high risk, transparency obligation, minimal risk) and identify your role: provider or deployer.
- Check the obligations that already apply: prohibited practices, Article 50 transparency, staff skills.
- Start the long-running work now for potentially high-risk systems: risk management, data quality, logging, human oversight and cybersecurity.
- Include AI-specific threats in your risk analysis and your penetration tests.
- Set rules for your suppliers of AI models and services: data location, use for training, incident notification.
- Formalize governance: acceptable use policy, AI lead, regular review of uses.