What matters in AI.

Subscribe

Learn / AI security

Definition · AI security

Confidential computing

Confidential computing is the protection of data in use by computation in a hardware-based, attested Trusted Execution Environment (TEE). The TEE is an isolated environment that prevents unauthorized access to code and data in use. It reduces the trust you must give the cloud operator. For AI, it can protect prompts, user data and models.

Last reviewed

Key points

  • Encryption at rest and in transit leaves one weak point. Data must be decrypted for use. Confidential computing closes that point with a hardware-based TEE.
  • Attestation makes it useful. A remote party checks signed evidence about the TEE before it sends a key or a prompt.
  • NVIDIA called the H100 the first GPU with confidential computing (August 2023). AI inference can run in a TEE on the CPU and the GPU.
  • Meta (WhatsApp) uses a TEE design for AI on user messages. Apple uses attested nodes for cloud AI. Anthropic published a research sketch of confidential inference.
  • The TEE does not stop physical attacks. TEE.fail (IEEE S&P 2026) took secret keys from Intel and AMD servers with a memory bus device that costs less than $1,000.

How it works

Data has three conditions. Confidential computing protects the one that ordinary encryption leaves open.

Condition of data Usual protection Weak point
At rest (disk) Disk encryption None while stored
In transit (network) TLS None on the network
In use (memory, processor) Confidential computing Must be decrypted for use

The processor makes a Trusted Execution Environment (TEE). The TEE is an isolated, hardware-enforced area for code and data. Examples are Intel TDX and SGX, AMD SEV-SNP, Arm TrustZone and the NVIDIA H100 GPU in confidential mode.

Attestation makes the TEE useful to a remote user. IETF RFC 9334 names three roles:

  1. Attester. The TEE sends signed evidence of what hardware and code it runs.
  2. Verifier. A service checks the evidence and gives a result.
  3. Relying party. The data owner releases a key or a prompt only if the result is good.
A data owner sends a secret only after a verifier accepts the TEE's evidence.
A data owner sends a secret only after a verifier accepts the TEE's evidence.

Why AI needs it

A cloud AI service must see the prompt in clear text to give an answer. Apple says that cloud AI needs unencrypted access to the request and the personal data of the user. Confidential computing limits who can see that text on the host.

It protects two assets:

  • User data. Prompts, files and outputs stay inside the TEE. This limits LLM data leakage to the operator.
  • Model weights. NVIDIA calls a trained model “highly valuable intellectual property”. In the Anthropic research sketch, model weights stay encrypted until the loader decrypts them.

In practice

Four public designs for AI, with dates:

Design Owner Date What it uses
H100 confidential mode NVIDIA Aug 2023 GPU TEE, encrypted CPU to GPU bounce buffer
Private Cloud Compute Apple Jun 2024 Attested nodes, public transparency log
Private Processing Meta (WhatsApp) Apr 2025 CPU confidential VM and confidential mode GPUs
Confidential Inference Anthropic Jun 2025 Trusted VM, TPM-based attestation (research sketch)

An Apple device encrypts each request only to nodes with attested software that is in a public log. Apple does not use the term confidential computing for this design.

Anthropic says “Not all accelerators fully support confidential computing yet”. Its design is a research sketch.

Trade-offs

Cost in speed. NVIDIA measured the encrypted CPU to GPU path at about 4 GB each second (August 2023). Zhu and colleagues measured Hopper GPUs in 2024. The overhead was below 7% for most typical LLM queries.

Physical attacks are out of scope. TEE.fail (IEEE S&P 2026) put a DDR5 bus device on Intel and AMD servers. The researchers say it costs less than $1,000 to make.

  • They took the Intel attestation key and forged TDX and SGX quotes.
  • They took private signing keys from an AMD SEV-SNP virtual machine.
  • They say NVIDIA GPU attestations “are not bound to the identity of a specific CVM”.

Intel and AMD say these attacks are out of their threat model. For the Intel TDX attack, the TEE.fail authors say there is “no current mitigation besides separately ensuring physical security of the TDX-enabled server”.

Questions and answers

Does confidential computing stop physical attacks on the server?

No. Intel and AMD put physical memory interposition out of the threat model of TDX, SGX and SEV-SNP. The TEE.fail researchers took the Intel attestation key with a DDR5 bus device. They say it costs less than $1,000 to make. For the Intel TDX attack, they say there is "no current mitigation besides separately ensuring physical security of the TDX-enabled server".

How is confidential computing different from encryption at rest and in transit?

Encryption at rest protects stored data. Encryption in transit protects data on the network. The two stop when software must use the data. Confidential computing protects the third condition, data in use. It keeps the data in an isolated, attested hardware TEE.

Can confidential computing protect AI model weights?

It can help, with limits. NVIDIA says confidential computing can protect both AI models and data. The Anthropic research sketch stores weights encrypted and decrypts them at the loader. It is not a product. The protection depends on attestation and on the physical security of the server.

Sources

  1. Azure Confidential Computing OverviewMicrosoft Learn, 30 Jun 2026
  2. Confidential Computing ConsortiumConfidential Computing Consortium (Linux Foundation)
  3. RFC 9334: Remote ATtestation procedureS (RATS) ArchitectureIETF, Jan 2023
  4. Confidential Computing on NVIDIA H100 GPUs for Secure and Trustworthy AINVIDIA, 3 Aug 2023
  5. Confidential Computing on NVIDIA Hopper GPUs: A Performance Benchmark StudyarXiv (Zhu et al.), 6 Sep 2024
  6. Private Cloud Compute: A new frontier for AI privacy in the cloudApple Security Research, 10 Jun 2024
  7. Building Private Processing for AI tools on WhatsAppMeta Engineering, 29 Apr 2025
  8. Confidential Inference via Trusted Virtual MachinesAnthropic, 18 Jun 2025
  9. TEE.fail: Breaking Trusted Execution Environments via DDR5 Memory Bus InterpositionGeorgia Tech, Purdue University and Synkhronix (research site), 28 Oct 2025

In the news