· Updated
ML-KEM-768 (a NIST-standardized post-quantum key encapsulation), HKDF-SHA256, and ChaCha20-Poly1305, then sent as ciphertext through Chutes' API and load balancers. Only the GPU instance inside a Trusted Execution Environment (TEE) sees the plaintext. A fresh ephemeral keypair per request provides forward secrecy.This shifts AI privacy from "trust the provider" to mathematically excluding the provider from the trust chain. For teams handling sensitive data — legal, medical, or financial — inference can run on external infrastructure without exposing prompt content. Available across all models today, with strongest guarantees on TEE-enabled models.
Teams using the OpenAI Python SDK can activate encryption via Chutes' chutes-e2ee transport. Teams on other platforms can run the e2ee-proxy Docker container, which supports both OpenAI-compatible APIs and Anthropic's Messages API. Both are MIT-licensed open source.
Most AI providers ask you to trust them with your data. We just removed ourselves from the equation. Today we're shipping end-to-end encryption for AI inference on Chutes. Here's what that actually means: https://t.co/u6iWo0ZaoC
More like this
CloudflareApr 27
ClaudeYesterday
Fireworks AIApr 30
AnthropicMay 4