One year later: Sovereign AI and the fight for choice
Source
blog.cloudflare.com
Date
Context
Cloudflare argues that an organization's AI defenses should remain usable when access to a model provider changes. Its sovereignty update describes an open-source coordination layer that assigns vulnerability hunting, verification and threat prioritization to multiple models.
The intended benefit is the ability to substitute models within the security workflow. The company says it has used open and closed models together and is launching workshops where government cyber agencies and infrastructure operators build and adapt their own systems. This frames model choice as an operational design requirement, with organizations retaining control over which models perform the work.
Key takeaways · AI-distilled
EuroLLM covers 35 languages, including all 24 official EU languages, and was built by a consortium including IST, Edinburgh and Paris-Saclay on the MareNostrum 5 supercomputer. The consortium says it outperforms similar-sized models on EU multilingual benchmarks and translation.
Apertus publishes its architecture, weights, training data and methods. ETH Zurich, EPFL and CSCS trained it on 15T+ tokenThe chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.Full definition → in 1,500+ languages, 40% non-English, respecting training opt-outs and removing personal data with the EU AI Act and GDPR in mind.
Cloudflare's security team built and open-sourced a agent harnessThe scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.Full definition → that runs multiple AI models in parallel to hunt vulnerabilities, verify findings and prioritize threats. Because it works with any closed or open model, losing access to one provider does not switch the defenses off.
Cloudflare is starting hands-on workshops where government cyber agencies and critical infrastructure operators build their own AI security harness and layered defenses. The first runs at Singapore International Cyber Week in October.
Terms in this piece · Glossary
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
