Why does attesting my LLM inference still let the operator infer what I asked?

Trusted Execution Environments can now prove a specific model ran without modification, which is real progress on integrity. But attestation proves integrity, not confidentiality. During transformer inference the memory access patterns, cache timing, and PCIe bus traffic between the CPU and GPU create a measurable side channel that leaks information about the prompt even inside an attested enclave. September 2026 research demonstrates verbatim token leakage through this channel on H100 configurations. Splitting the model across a trusted CPU and an untrusted GPU, which is required at production scale, enlarges the attack surface in ways that current enclave designs do not address.