Threat Intelligence

Tens of Thousands of Relay Servers Found Masking Access to US AI Models

Dark Reading · 23 Sept 2026

Key Takeaway If your business relies on geographic or identity-based access controls to protect valuable data or tools, don't assume they are foolproof, monitor for unusual usage patterns as an added layer of defence.

Security researchers have uncovered a network of more than 80,000 relay servers that appear to be helping users mask their true location while accessing advanced large language models (LLMs) built by US companies. The relays make it look as though requests are coming from elsewhere, hiding the fact that many users are believed to be based in China.

The likely motive, according to researchers, is to clone or replicate the capabilities of frontier AI models without directly identifying who is accessing them. This kind of activity raises concerns for AI providers about intellectual property theft and unauthorised use of their systems, and highlights how relay and proxy infrastructure can be used to evade access controls and geographic restrictions.

While this case centres on AI companies rather than typical small businesses, it is a reminder that any organisation offering online services, including APIs, customer portals or paid tools, should be aware that access restrictions based on location or identity can be bypassed using relay networks. Monitoring for unusual access patterns remains an important defensive measure.

AI security data theft relay networks threat intelligence access control

Summarised by CISO AI from Dark Reading. We link back to every original so you can read it yourself.