AI and cybersecurity: the overlap of the Cyber Resilience Act and the AI Act

Short answer: An AI product must not only function safely but also withstand attacks. The Cyber Resilience Act (CRA, Regulation (EU) 2024/2847) sets cybersecurity requirements for virtually all products with digital elements. The AI Act additionally requires an appropriate level of cybersecurity for high-risk AI systems (Article 15). Both frameworks can apply to the same product.

What the Cyber Resilience Act governs

The CRA is a horizontal framework: it applies to products with digital elements, from software to connected devices. Core obligations are:

• Security by design: products are designed and supplied without known exploitable vulnerabilities.
• Vulnerability handling: manufacturers provide security updates during a support period.
• Reporting duty: actively exploited vulnerabilities and serious incidents must be reported.

Many AI systems are supplied as or within such products and thus fall under the CRA.

The requirements apply throughout the lifecycle, not only at first sale. A manufacturer must keep monitoring vulnerabilities and roll out patches for as long as the product is supported. This reflects the idea that security is a continuous obligation, not a one-off test.

What the AI Act requires

The AI Act approaches cybersecurity from the risk of the AI system itself. Article 15 requires that high-risk systems have an appropriate level of accuracy, robustness and cybersecurity.

This includes AI-specific threats: data poisoning (corrupting training data), adversarial attacks (deceptive input) and model evasion. These are risks that a classic security approach does not automatically cover.

An attacker need not crack the software to undermine an AI system; manipulating input or training data can be enough to force wrong outcomes. That is why the AI Act requires robustness and cybersecurity to be built into the design from the start.

One product, two conformity tracks

The big challenge is overlap without duplicated work. The CRA and the AI Act both pursue safe products, but from different starting points: the CRA generic, the AI Act AI-specific. The legislator intended alignment, so that a product subject to both does not have to run two entirely separate processes.

In practice this means: one risk assessment covering both angles, linked to the right transparency and documentation. As with sustainability reporting, it pays to bundle frameworks rather than run them in parallel.

What to do

• Map both frameworks: does your product fall under the CRA, the AI Act, or both?
• Apply security by design: ship without known exploitable vulnerabilities.
• Address AI-specific threats: test for data poisoning and adversarial attacks.
• Set up vulnerability and reporting processes in line with the CRA.
• Integrate the risk assessments of the CRA and AI Act into one coherent process.

Cyber resilience is not a separate layer on top of AI, but part of the robustness that both frameworks require.