The EU AI Act in plain English
The EU AI Act (Regulation (EU) 2024/1689) is considered the first comprehensive law to regulate artificial intelligence (AI). As an EU regulation, it applies directly in all member states, including Germany. And it doesn’t only affect companies that develop AI. Anyone who buys and uses AI – such as a customer service chatbot or software that pre-screens job applications – has obligations, too.
The basic principle is familiar from road traffic: a bicycle is subject to just a few rules, a truck carrying hazardous materials to a great many. The AI Act follows the same logic and scales its rules to the risk. The greater the potential harm to health, safety, or fundamental rights, the stricter the requirements.
One point is particularly important for IT security: the regulation explicitly requires high-risk AI systems to be protected against attacks. That makes AI a topic not only for data protection and ethics, but also for security testing.
What the EU AI Act regulates
The regulation sorts AI applications into risk levels:
- Prohibited practices: AI that deliberately manipulates people or exploits their vulnerabilities in ways that can cause them significant harm is banned. Also prohibited are social scoring – rating people based on their social behavior – and, as a rule, emotion recognition in the workplace.
- High-risk AI: This includes AI systems in sensitive areas such as recruitment, credit scoring, education, or critical infrastructure. It also covers AI used as a safety component of regulated products, such as medical devices. Strict requirements apply here: risk management, data quality, documentation, and human oversight.
- Transparency obligations: People must be able to tell that they are talking to a chatbot. Deepfakes – deceptively realistic, artificially generated images, videos, or voices – must be disclosed as such.
- Minimal risk: Spam filters or AI in video games are not subject to any specific obligations.
Separate rules apply to general-purpose AI models, such as large language models (LLMs), which many applications are built on.
For high-risk AI, Article 15 requires an appropriate level of accuracy, robustness, and cybersecurity. Robust means that the system works reliably even when errors and disruptions occur. Cybersecurity here means that it withstands attacks – such as manipulated training data (“data poisoning”) or crafted inputs designed to trick the model into making mistakes.
The obligations apply in stages:
- since February 2, 2025: prohibitions and measures to build AI literacy among employees
- since August 2, 2025: rules for general-purpose AI models
- since August 2, 2026: further obligations, including the transparency obligations
- from December 2, 2026: additional bans on AI that generates intimate images without consent or child sexual abuse material
- from December 2, 2027: high-risk AI in sensitive areas of use
- from August 2, 2028: high-risk AI in regulated products
The additional bans and the later high-risk deadlines were introduced by an amendment from July 2026, the “Digital Omnibus” on AI (Regulation (EU) 2026/1744). That’s why older sources give different dates.
A practical example
Here’s a typical scenario: A retail chain is constantly hiring new staff. It introduces AI software that pre-screens applications and ranks the candidates. Systems like this generally count as high-risk AI. The company acts as the “deployer”: it must use the system in accordance with the provider’s instructions, put human oversight in place, and monitor its operation.
Even before these obligations take effect in December 2027, the company has the application – including its chat feature for applicants – tested for security vulnerabilities. The result: with cleverly worded inputs, known as prompt injection, the chat reveals internal evaluation criteria and excerpts from other people’s applications. The provider then restricts which data the chat can access before the system goes live.
Why does this matter for businesses?
The AI Act also affects companies that “only” use AI – which obligations apply depends on their role. Providers place an AI system on the market under their own name and have usually developed it themselves. Deployers use it under their own responsibility. Violations can be expensive: for prohibited practices, fines can reach €35 million or 7% of worldwide annual turnover, whichever is higher. For small and medium-sized enterprises, the lower amount applies in each case.
Regardless of the law, security is in your own interest. AI applications are often connected to internal documents or databases. That creates new risks, such as confidential data leaking out. And if the AI processes personal data, the GDPR applies as well.
What businesses should do now
- Take inventory: Record which AI systems you use or develop – including those built into purchased software and cloud services.
- Clarify your role: For each system, determine whether you are the provider or the deployer.
- Classify the risk: With legal support, assess whether an application could be prohibited or high-risk.
- Build AI literacy: Train employees and set rules for using AI tools.
- Create transparency: Label chatbots and AI-generated content.
- Test security: Check AI applications for prompt injection, data leakage, and misuse of connected functions before go-live and after major changes.
- Document: Keep a traceable record of decisions, tests, and measures.
How Rootsektor can help
In our AI & LLM penetration test, we check chatbots, AI assistants, and AI agents for prompt injection, data leakage, and unintended actions. Our testing is guided by the OWASP Top 10 for LLM Applications. The report shows which attacks your system withstands – a technical building block for demonstrating robustness and cybersecurity, but not a legal classification. You can get a fixed-price quote after scoping via our Pentest Quote page.