Confidential AI for GitLab Self-Hosted

This post details the integration of GitLab Duo Self-Hosted with Privatemode AI, enabling the use of confidential computing for AI coding agents. It explains how Privatemode's hardware-based trusted execution environments (TEEs) and remote attestation keep prompts and source code encrypted during inference, addressing the needs of regulated organizations. The integration involves running a self-hosted AI Gateway and a Privatemode proxy, which exposes an OpenAI-compatible API. The post outlines the architectural answer of confidential computing, the tradeoffs involved (operating the gateway/proxy, maturing generic models, attestation overhead), and the security properties of the solution.