Trusted execution environments allow for end-to-end encryption of AI prompts and outputs during inference