Securing a European Sovereign Inference Provider: a Technical Case Study with Xference and HikmaAI

About this session

European companies are now running LLMs in production. Inference providers are the plumbing underneath, which means they are critical infrastructure, and they present an attack surface nobody really knows how to defend. This talk covers a joint technical case study between Xference, a Sovereign European inference provider, and HikmaAI, showing how we designed a two-stage security architecture, adversarial testing before deployment, and runtime enforcement via the HikmaAI Intelligence Gateway. We walk through the threat model, the validation methodology we used against models and agentic workflows, the attack classes we caught before going live, and the runtime controls that now block prompt injection, jailbreaks, data exfiltration attempts, and policy violations on live traffic. We share real numbers, attacks found, attacks blocked, latency overhead, false positive rates, and without killing private inference performance.

Speakers

Andrea Pili

CEO & Founder, Xference




Mauro Medda

CTO, HikmaAI