How confidential AI splits control between data and model owners — and opens new opportunities for both
The New Stack Megan Carnegie ● Covered by 2 sources
Enterprises want AI on sensitive data without handing it to the cloud. Confidential AI keeps the model builder’s weights and the customer’s data under separate control.
Based on reporting by The New Stack, Megan Carnegie — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Enterprises like the idea of generative AI a lot more than they like the plumbing problem it creates. If a model needs patient records, financial data, or company IP, sending that information to a cloud service means moving it across outside networks and onto someone else’s infrastructure. That’s where the neat demo meets the ugly part: trust, control, and accountability.
Alon Horev, CTO and co-founder of VAST Data, says the pain is sharpest when customer information is involved. Even something as basic as redacting a conversation or stripping out PII is hard to guarantee with total confidence. And once a support agent or other system touches that data, the organization has to worry about exposure to the wrong customer, plus whether the conversation can be reused for training or improvement.
The core problem is that both sides want control. Enterprises want their data to stay theirs. Model builders want to protect weights and software that took serious research and engineering to create. SaaS delivery helps the model side distribute and defend its product, but it also pushes customers into a setup where more systems, operators, and links have to be trusted.
Confidential AI is the answer VAST Data is pushing with VAST DataEnclave, launched on September 22. The idea is to use encrypted processing, hardware-isolated execution, and cryptographic attestation so data and model weights can stay encrypted until an approved workload is cleared. The customer keeps its keys for data. The model builder keeps its keys for the model. Nobody in the middle gets both.
That arrangement matters even more as agents become more capable and less supervised. Horev’s view is blunt: if today’s systems can’t prove they’re safe with sensitive data, agents will be even harder to deploy safely. VAST’s pitch is that the stack has to be built as one living system, with storage, networking, compute, observability, and access controls all working together rather than bolted on after the fact.
My take — AI-written commentary, not fact-checked reporting
This is the part of AI everyone skips when they’re busy admiring the demo: the model is useless if the governance story is hand-wavy. Confidential AI is basically the industry admitting that “just trust the cloud” is not a policy. That’s healthy, if a little embarrassing for the last few years of hype.
Read more about this at: The New Stack