Umbra blog

Private, low-cost inference — and the models nobody else hosts.

How verifiable prompt privacy works, why it runs on idle Apple-Silicon Macs, and field notes from the alpha.

Jun 27, 2026How Apple attestation + AMD SEV-SNP prove your AI prompts are privateMost 'private' LLM APIs just promise not to log you. Umbra lets you cryptographically verify your prompt was never readable — by the host or the operator. Here's the attestation chain, end to end.