What is Blackthorn?
The problem it solves and what you get.
How it works
How trust is split across GPU Confidential Computing and Arcium’s MPC network.
Try it
Quickstart
Clone, build and run the stack on your own machine — no GPU required.
INTRODUCTION
Blackthorn
Arcium’s architecture for confidential LLM inference on NVIDIA Confidential Computing GPUs, with no CPU TEE in the trust chain.
Arcium Blackthorn lets you use an AI model running on someone else’s servers without the people running those servers ever seeing your prompts, your files or the outputs. Your data is encrypted on your own device and only decrypted inside the GPU’s protected memory.
Confidential AI has traditionally been built by pairing the GPU with a CPU TEE, which asks you to trust that CPU with the GPU’s session keys, the keys that protect your data.
Those CPU TEEs have been broken repeatedly, and a single break can expose every prompt, output and model weight behind them, for every user on the machine. Blackthorn is a new approach that changes the stack, removing the CPU TEE from the trust chain entirely and using MPC in its place alongside NVIDIA Confidential Computing on the GPU.
This means no publicly documented attack on CPU TEEs applies to Blackthorn, including the physical-access attacks today’s confidential AI setups are exposed to through their CPU TEE.