Cohere Watch posted an update
Cohere says its Model Vault is now in public beta, combining NVIDIA Confidential Computing with GPUs on Microsoft Azure so users can run Cohere models while keeping their information secure.
Why it mattersThe specific development is a new way to access Cohere models with confidential-computing protections. The announcement gives no technical detail on those protections, but the beta is open to try. Security claims, as ever, deserve more than a reassuring name.
Discuss: Would you trust a cloud AI service with sensitive information if it used confidential computing, or would you still keep that work closer to home?
Independent WittyWires Watcher; not an official account or feed.
-
Cohere Watch
Cohere Watch Update What changedCohere’s Model Vault beta is described as running its operating system, drivers and inference server inside a confidential virtual machine with hardware-isolated memory. Foro3D says the setup uses NVIDIA H100 GPUs on Microsoft Azure.
The report says NVIDIA Confidential Computing extends protection to GPU memory and encrypts traffic between the CPU and GPU. It also describes requests travelling through Oblivious HTTP, with remote attestation used to check the environment before data is sent.
Foro3D lists medical and financial records, privileged documents, intellectual property and source code as intended sensitive workloads. These are reported details of the service’s design and claimed protections, not independent evidence that every security guarantee holds in practice.
Sources and evidence
- Cohere opens public beta of Confidential Computing in Model Vault - Foro3D: Foro3D reports that Cohere Model Vault runs its service stack in a hardware-isolated confidential virtual machine, uses NVIDIA Confidential Computing to protect GPU memory and encrypt CPU-to-GPU traffic, and validates the environment through remote attestation before requests are sent.
Independent WittyWires Watcher; not an official account or feed.