How the CRA Applies to AI and ML Products
Does the Cyber Resilience Act apply to AI and ML products? Discover the scope, specific obligations, and interaction with the EU AI Act.
Does the Cyber Resilience Act apply to AI and ML products? Discover the scope, specific obligations, and interaction with the EU AI Act.
Artificial intelligence is transforming digital products, but where is the boundary between the Cyber Resilience Act and the EU AI Act? If your product incorporates AI or ML components, here is how to navigate these two major regulations.
Yes, if the AI product is a product with digital elements under the CRA. This includes software applications embedding ML models, IoT devices using on-device AI, SaaS platforms with generative AI features, and commercialized LLM APIs. However, an AI model distributed solely as a service (via API) without being integrated into a product may fall under different frameworks.
The EU AI Act and the CRA are complementary but distinct: the AI Act regulates risks from the use of AI itself (transparency, human oversight, algorithmic bias), while the CRA regulates cybersecurity of the digital product that embeds or uses AI. An AI product may be subject to both regulations simultaneously.
The CRA requires products to be resilient against data corruption attempts. For AI products, this means protecting training datasets from poisoning, validating data pipeline integrity, and logging training data access.
Inference and model extraction attacks must be addressed: protection against model stealing, adversarial attacks (modified inputs to deceive the model), and abnormal API usage detection.
The SBOM must include AI frameworks, libraries, and models used, with their respective versions and licenses. Pre-trained models must be traced with their source and exact version.
The boundary between product security (CRA) and AI ethics (AI Act) is sometimes blurred, but compliance with both is essential for bringing innovative AI products to the European market.