Did the Cloudflare AI incident affect all Cloudflare customers?

No. According to Cloudflare's post-mortem of 26 October 2026, it affected one model (GLM-5.3) on Workers AI, only requests with very long prompts, and 155 responses between 27 August and 9 September 2026. Cloudflare found no evidence of deliberate exploitation.

Is shared cloud AI unsafe?

Not in general. Providers isolate customers in software, and incidents are rare. The risk is that your data's safety depends on someone else's software. For public or test data that is usually acceptable; for personal, confidential or regulated data, private AI on your own servers removes that dependency.

What is on-premise or private AI?

AI that runs on servers you control, in your own data center or a dedicated environment, with no connection to an outside AI service. Your documents and prompts never leave your network.

Who builds private on-premise AI in Uzbekistan?

Eurosoft, a software and AI company in Tashkent and a resident of IT Park, builds private LLM and RAG platforms that run fully on the client's servers. Its Enterprise RAG Platform (SarhadAI in Uzbekistan) runs on-premise at Samsung and Incheon International Airport.

Shared cloud AI or private AI: how to keep company data out of other people's answers

What happened

Between 27 August and 9 September 2026, a small number of customers using the GLM-5.3 model on Cloudflare Workers AI received answers that contained data from other customers' requests to the same model.
The cause was a combination of a configuration mismatch and a bug in how GPU memory (the KV cache) was reused between requests. Only very long prompts were affected, 155 responses in total, and Cloudflare found no evidence of deliberate exploitation.
Cloudflare took the model offline, fixed the configuration, patched the open-source inference engine and published the details on 26 October 2026.

Why shared AI can mix up data

On a shared AI platform, many companies send requests to the same model on the same GPUs. To be fast and cheap, the platform reuses memory between requests.
Keeping each company's data apart is the provider's job, done in software. When that software has a bug, data can cross from one customer to another, even when nobody is attacking anyone.
This is not a reason to avoid cloud AI. It is a reason to decide, for each kind of data, whether you are comfortable relying on someone else's isolation.

When shared cloud AI is fine

Public information: marketing texts, translations of published material, general research.
Prototypes and pilots where you test an idea on sample data.
Tasks where speed of launch and low cost matter more than full control.

When you need private AI

Personal data of citizens, patients or students. In Uzbekistan the personal data law requires citizens' personal data to be stored on servers in Uzbekistan.
Confidential company documents: contracts, financial reports, internal procedures, designs and source code.
Regulated sectors: government, banking, healthcare, telecoms, energy and transport.
Cases where you need to prove where data lives and who can see it, for an auditor, a regulator or a large client.

What private AI looks like in practice

Private AI means the model and your documents run on servers you control: in your own data center or in a dedicated environment, with no connection to an outside AI service.
Eurosoft's Enterprise RAG Platform works this way at Samsung and Incheon International Airport: it runs on the client's own GPUs, including a 120B-parameter open model, fully offline, with no external API. Every department has its own knowledge base, and an access check runs before every search. In Uzbekistan the platform is offered as SarhadAI.
Open models such as GPT-OSS, Llama, Qwen and Gemma now make this possible without sending data to a model provider.

Questions to ask your AI vendor

Where exactly do my prompts and documents run: which country, which data center, shared or dedicated hardware?
Is my data stored, logged or used to train models? For how long?
How is my data kept apart from other customers' data, and how is that tested?
Can the system run fully on our servers, without internet access?
Who can see answers and documents inside our company? Is access checked per department?

Questions and answers

Did the Cloudflare AI incident affect all Cloudflare customers?: No. According to Cloudflare's post-mortem of 26 October 2026, it affected one model (GLM-5.3) on Workers AI, only requests with very long prompts, and 155 responses between 27 August and 9 September 2026. Cloudflare found no evidence of deliberate exploitation.
Is shared cloud AI unsafe?: Not in general. Providers isolate customers in software, and incidents are rare. The risk is that your data's safety depends on someone else's software. For public or test data that is usually acceptable; for personal, confidential or regulated data, private AI on your own servers removes that dependency.
What is on-premise or private AI?: AI that runs on servers you control, in your own data center or a dedicated environment, with no connection to an outside AI service. Your documents and prompts never leave your network.
Who builds private on-premise AI in Uzbekistan?: Eurosoft, a software and AI company in Tashkent and a resident of IT Park, builds private LLM and RAG platforms that run fully on the client's servers. Its Enterprise RAG Platform (SarhadAI in Uzbekistan) runs on-premise at Samsung and Incheon International Airport.