Open Anonymity Project Launches Privacy-Focused Access to AI Models

Summary

The Open Anonymity Project presents an AI assistant designed to let users access open and closed models without linking queries to their identity. Its unlinkable inference protocol separates a user’s requests from one another and from the user’s identity, so OA and model providers do not know which queries belong to that person. OA says its public verification code runs in a Trusted Execution Environment and can be used to verify that queries are not logged. The service also offers Parallel and Council modes, prompt scrubbing, and personal memory agents. A query rewriter, implemented with a small language model in a GPU enclave, is intended to conceal personally identifiable information and distinctive writing styles in prompts. Conversation history remains on the user’s device, while memory agents in secure GPU enclaves select relevant context without sharing the complete history. OA also describes an in-browser proxy developed with the Refraction Network to reduce revealing network metadata. For account access, users can sign up with a username or Google, use a passkey, or continue with an Ethereum wallet. The service integrates zkAPI, a zero-knowledge protocol for paid APIs including AI inference, which supports anonymous Ethereum payments and is described as a collaboration between the Ethereum Foundation and OA.