Leo Adds E2E Encryption, Oblivious HTTP, and Verifiable Confidential Computing (brave.com)

Brave has announced significant upgrades to its AI assistant Leo, introducing end-to-end encryption, Oblivious HTTP, and verifiable confidential computing to enhance user privacy. Conversations using Leo, powered by the open-weight GLM model, are encrypted on the user’s device before they leave, ensuring only a secure, attested Trusted Execution Environment (TEE) can decrypt them for processing. This architecture not only protects user data from being accessed by Brave or the AI provider, NEAR AI, but also offers cryptographic proofs of authenticity, helping to build trust in an era where data privacy is increasingly compromised. These innovations hold substantial implications for the AI and machine learning community, emphasizing the urgent need for privacy-conscious architectures. With the ability to confirm that only authorized hardware is handling sensitive data, Brave’s approach counters common risks associated with AI conversations, such as data leaks and unauthorized usage for model training. This shift not only safeguards personal information but also sets a precedent for future developments in AI privacy, pushing towards an ecosystem where openness, transparency, and verifiability are foundational principles, marking a notable evolution in user-first AI.