Overview: Cloudflare’s Push for Sovereign AI and Choice
Cloudflare emphasizes “Sovereign AI” as a strategic imperative, advocating for diversified, accessible AI capabilities to counter vendor lock-in and enhance national resilience. The company recently announced the integration of two new European open multilingual models, EuroLLM and Apertus, into its Workers AI platform. These initiatives aim to provide greater choice and flexibility for organizations seeking to develop and deploy AI solutions while maintaining control over their data and infrastructure, as detailed in their recent post, “One year later: Sovereign AI and the fight for choice”. This strategy directly addresses growing concerns about AI access restrictions and the potential misuse of frontier models in cyberattacks.
Technical Analysis of Sovereign AI Implementation
The concept of Sovereign AI centers on a nation’s ability to develop, control, and deploy AI technologies without undue reliance on external entities or single vendors. Cloudflare’s approach focuses on providing infrastructure that supports a multitude of models, fostering an environment where nations and organizations can select tools that align with their specific regulatory and security requirements.
New Open Multilingual Models on Workers AI
Two key additions exemplify this strategy:
- EuroLLM: Developed by a consortium including Instituto Superior Técnico, University of Edinburgh, and others, with support from Horizon Europe, EuroLLM covers 35 languages, including all 24 official EU languages. Trained on the MareNostrum 5 supercomputer, this model is specifically designed to serve underserved European languages and reportedly outperforms similar-sized models on EU multilingual benchmarks. This represents a significant step towards Sovereign AI implementation for national resilience within the European Union.
- Apertus: Switzerland’s first large-scale, fully open, multilingual language model, Apertus was developed by ETH Zurich, EPFL, and the Swiss National Supercomputing Centre (CSCS). Trained on over 15 trillion tokens across more than 1,500 languages (40% non-English), Apertus prioritizes compliance with European regulations like the EU AI Act and GDPR. Its architecture, weights, training data, and methods are all publicly available, fostering transparency and trust. Developers report its strong performance on rare and regional languages, offering crucial capabilities for diverse populations.
Developing Model-Agnostic AI Defenses
Cloudflare also highlights the critical need for building AI defenses with open models that are not dependent on a single provider. The company’s security team has developed AI defenses that function irrespective of the underlying model, ensuring that loss of access to one specific frontier model does not compromise an organization’s defensive posture. This proactive measure is particularly relevant given that attackers have already utilized frontier models to launch cyberattacks, underscoring the dual-use nature of advanced AI. Cloudflare plans to host hands-on workshops for government cyber agencies and critical infrastructure operators, starting in Singapore, to facilitate the development of these essential AI defense capabilities.
Actionable Recommendations and Mitigations
Organizations and national entities should consider the following actions to enhance their AI security and sovereignty:
- Prioritize Model Diversity: Avoid exclusive reliance on a single AI model or provider. Explore and integrate a range of open-source and specialized models, such as those now available on Cloudflare Workers AI, to reduce vendor lock-in and improve resilience.
- Invest in Model-Agnostic AI Defenses: Develop or adopt security solutions that can protect against AI-driven threats regardless of the specific large language model (LLM) being used or targeted. This ensures continuity of defense even if access to certain frontier models becomes restricted or compromised.
- Support Open AI Initiatives: Contribute to and leverage public research and open-source projects like EuroLLM and Apertus. These collaborations foster innovation, transparency, and collective security in the AI ecosystem.
- Engage in Training and Workshops: Participate in programs designed to educate security professionals on Cloudflare Workers AI new European models and how to deploy effective AI defenses, as offered by Cloudflare and similar organizations.
Related: AI Guardrails: Hindering SOCs and Aiding Adversaries, Context Bombing: Defending Against AI Hacking Agents