Скачать или смотреть Real-World AI Case Study: How Slack does AI Security

Real-World AI Case Study: How Slack does AI Security

A real-world case study breaking down how Slack built AI features for millions of enterprise users without letting customer data leave their infrastructure. Covers the trust boundary architecture, ACL-scoped data retrieval patterns, and the complete system design that enables secure RAG at enterprise scale.

Reference blog post by Slack: https://slack.engineering/how-we-buil...

================
What you will learn:
================
The Enterprise AI Problem
The fundamental tension between AI capabilities and data sovereignty
Why standard LLM integration breaks FedRAMP, HIPAA, and SOC 2 compliance
How Slack refused the "AI or compliance" trade-off

The Four Security Principles
Customer data never leaves the trust boundary
No training on customer data (eliminates fine-tuning risks)
AI operates only on user-visible data (permission alignment)
All compliance extends to AI features automatically

Trust Boundary Architecture
Escrow VPC arrangement for hosting models inside your infrastructure
How the model provider licenses weights without accessing customer data
Why customer data never exits Slack's infrastructure

RAG vs Fine-Tuning Decision
Why fine-tuning creates cross-tenant data leakage risk
How RAG's stateless design provides strong privacy guarantees
The architectural benefit of true statelessness

ACL Enforcement Pattern
Enforcing permissions at data retrieval, not output filtering
The key insight: ACL is passed into the query, not checked after
Reusing existing infrastructure for identical permission behavior

Production Code Patterns
ACL-scoped data retrieval implementation
Complete channel summarization flow with 6 components
DLP-aware ephemeral outputs

Real-World System Architecture
Slack VPC vs AWS Escrow VPC design
Request flow from user click to ephemeral response
How every security control executes in correct order

===========
Timestamps:
===========
00:00 - Introduction: Slack's Secure Enterprise AI
00:24 - Step 1: Core Concepts Overview
01:05 - The Four Security Principles
02:07 - Trust Boundary Architecture
03:03 - RAG vs Fine-Tuning Decision
03:57 - ACL Enforcement Pattern
04:19 - Ephemeral & DLP-Aware Outputs
05:31 - Step 2: Code Examples
05:43 - ACL-Scoped Data Retrieval Pattern
06:33 - Complete Summarization Flow
07:49 - Step 3: Real-World Architecture
07:53 - System Overview: Trust Boundary Design
08:36 - Request Flow: Summarize Sequence
10:00 - Summary: Key Architecture Patterns

=========
About me:
=========
I'm Mukul Raina, a Senior Software Engineer and Tech Lead at Microsoft, with a Master's in Computer Science from the University of Oxford.

#SlackAI #EnterpriseAI #RAG #AIArchitecture #TrustBoundary #ProductionAI #AISecurity #DataSovereignty #SystemDesign #LLM #AIEngineering #CaseStudy