SYSTEM Cited by 1 source
Confidential Virtual Machine (CVM)¶
A Confidential Virtual Machine (CVM) is a hardware-isolated VM instance whose memory, register state, and (with confidential-GPU extensions) accelerator state are inaccessible to the host OS, hypervisor, and cloud operator. The CVM is the VM-granularity realisation of a Trusted Execution Environment (TEE): the guest runs a full Linux-style OS + application stack inside a hardware-enforced boundary, with memory encrypted by the CPU (e.g. via AMD SEV-SNP or Intel TDX) and execution isolation enforced outside the hypervisor's trust boundary.
Meta builds WhatsApp Private Processing on CPU-based CVMs plus Confidential-Compute-mode GPUs, explicitly so that LLM inference (the GPU-side work) is also inside the confidentiality boundary — not just the orchestration code.
Key properties¶
- Memory encrypted in hardware — the host cannot read the guest's pages; swap and DMA routes are CPU-gated.
- Hardware root of trust — the CPU produces an attestation of the loaded binary that a remote party can verify cryptographically, without trusting the host operator.
- Boot-time + runtime integrity — the TEE can be measured at boot and selected runtime points, so modifications to the loaded image are detectable.
- Confidential-Compute GPUs — modern datacentre GPUs expose a mode in which the GPU's memory + command queues are similarly isolated, so model weights + activations + outputs stay inside the boundary during inference.
- Reduced TCB — the host operating system + hypervisor are removed from the trusted-computing base (TCB); only the CPU vendor, the CVM's own software, and the attestation verifier need to be trusted.
What CVMs do NOT solve on their own¶
- Code trust — attestation proves which binary is running, not that the binary does what it claims. This is why CVMs are typically paired with a transparency log of acceptable binary digests + open-source + external review.
- Side-channel attacks — TEE software-exploitation research repeatedly demonstrates bypass capabilities; this is the reason Meta layers defence-in-depth on top of the CVM, not instead of it.
- Physical attacks — encrypted DRAM + multi-party review close many, not all; Meta explicitly names this residual risk.
- Correctness of the application inside the CVM — if the app has an exploitable bug, the CVM doesn't help. Containerisation, hardened binaries, log-filtering, minimised input surfaces all address this.
Meta's disclosed CVM surface¶
From the Private Processing 2025-04-30 post:
- CPU-based confidential virtualisation technologies — vendor not named (AMD SEV-SNP / Intel TDX / Arm CCA all plausible).
- Confidential Compute mode GPUs — GPU vendor + generation not named (NVIDIA Hopper CC is the obvious candidate as of 2025-04).
- Encrypted DRAM — called out explicitly.
- Remote shell access prohibited — including from the host.
- Attestation checked against a third-party ledger of acceptable binaries before client session-key release.
- CVM-to-CVM calls ride the same RA-TLS primitive clients use.
Seen in¶
- sources/2025-04-30-meta-building-private-processing-for-ai-tools-on-whatsapp — Meta's WhatsApp Private Processing uses a CVM + Confidential-Compute-mode GPU as the trust boundary for private AI inference. Canonical wiki instance.
Related¶
- concepts/trusted-execution-environment — the generic primitive class.
- concepts/confidential-computing — the industry-wide posture.
- concepts/remote-attestation — hardware-rooted binary-identity proof.
- concepts/ra-tls — the TLS composition that gates session keys on attestation.
- systems/whatsapp-private-processing — canonical wiki deployment.