How We Built Slack AI To Be Secure and Private

This post details the architecture of Slack AI, focusing on security and privacy. Key contributions include: 1. Hosting foundational LLMs within Slack's trust boundary using AWS escrow VPCs to prevent data leakage to model providers. 2. Employing Retrieval Augmented Generation (RAG) to ensure stateless AI operations, preventing LLM training on customer data and grounding results in company knowledge. 3. Enforcing that Slack AI only operates on user-visible data by reusing existing ACLs and data fetching libraries. 4. Integrating Slack AI with existing enterprise compliance infrastructure, such as Encryption Key Management and International Data Residency, and ensuring derived content (e.g., summaries) is invalidated if the source message is tombstoned by DLP.