AI Firms Pledge Unbreakable Data Privacy

Canadian AI startup Cohere announced a significant advancement in enterprise data privacy with its Model Vault product, now featuring confidential computing. This new capability assures businesses that their sensitive data remains private and inaccessible to the AI provider or cloud infrastructure. Traditional data protection methods like encryption at rest and in transit are enhanced by confidential computing, which utilizes hardware to encrypt and isolate data in memory during AI processing.

Cohere's solution ensures that prompts and data are protected the moment they are processed by AI models. This protection extends across both CPU and GPU components, a notable improvement over prior confidential computing deployments. The system provides a hardware boundary, with data remaining encrypted throughout its journey within the confidential environment and only decrypted for processing. Customers receive an attestation report to verify the security measures, with plans for open-sourcing the serving stack to allow independent auditing.

This development comes as major AI players like OpenAI and Anthropic face scrutiny over data handling practices. Companies like Nvidia and Booz Allen Hamilton have reportedly restricted their use of certain AI models due to data retention policy changes. Cohere's offering addresses this growing concern by providing a verifiable way to ensure data privacy, moving beyond vendor promises to hardware-backed assurances. The company aims to build trust through governability and security, positioning its technology as a secure choice for enterprises handling regulated or proprietary information.