Sweden's Cybersecurity Act is in force: three consequences for your AI work

Sweden missed the EU's NIS2 transposition deadline in October 2024 — then compressed twelve months of preparation into a couple of months. The Riksdag adopted the new Cybersecurity Act (SFS 2025:1506) in December 2025, and it has been in force since 15 January 2026. For organisations that are simultaneously building and using AI, this creates a rarely discussed intersection. Here are the three consequences we consider most concrete.

1. Your AI systems are part of the ICT estate that must be registered and governed

The Cybersecurity Act requires registration with NCSC/FRA and established cybersecurity governance for entities in scope. AI services are not exempt: an LLM integration reading internal documents, or an agent touching business systems, is part of the digital infrastructure that must be inventoried, risk-assessed and monitored.

Most AI inventories we see today lack exactly what the law requires: who owns the system, which data flows where, and how it connects to the existing incident process. That is the same inventory the EU AI Act demands — do it once, structured for both.

2. The incident clocks apply to AI-related events too

The law introduces reporting obligations with clocks at 24 hours, 72 hours and one month. An AI-related incident — documents leaked through a misconfigured RAG solution, a manipulated agent, customer data egressing to a model provider — is, from the law's perspective, an ICT security event to be assessed against the same clocks.

IBM's Cost of a Data Breach 2025 shows that one in five organisations has already experienced breaches linked to shadow AI, and that breaches in systems without AI access controls are significantly more expensive. In other words: the events are already happening. The question is whether your organisation gets them registered on the right track in time.

3. The supply chain is audited — for real this time

The law requires managing risks connected to direct suppliers and subcontractors. For AI specifically, the questions get sharper: where is the model hosted, are prompts and responses retained, and which terms apply at the model provider?

We recommend a single supplier assessment for AI — the same basis as the AI Act's supplier controls and your customers' own audits. One assessment, three purposes.

The way forward

The law's ambitions are high and the timeline was short — many entities in scope are currently making up lost ground. Our advice: start with the inventory (it is shared with the AI Act), connect AI events to the existing incident process, and treat model providers as the ICT suppliers they are.

Sources

• Government Offices of Sweden — New Cybersecurity Act (NIS2), adopted 15 December 2025, regeringen.se
• SFS 2025:1506, Cybersecurity Act, riksdagen.se / lagen.nu
• IBM — Cost of a Data Breach Report 2025, published 30 July 2025