BlueHat IL 2026 - Ido Shani, Gal Zaban - ChainLeak: From AI Framework to Cloud Secrets

AI Application Architecture

Gal Zaban and Ido Shani of Safran present ChainLeak, research on the Chainlit Python framework used to build enterprise AI applications. They map the stack: UI, backend agents and tools, orchestration layer, and local or cloud-hosted LLMs, noting rising vulnerabilities across this architecture.

Why Chainlit Matters

Chainlit is widely adopted with roughly one million monthly downloads and many internet-facing deployments. The speakers examine its callbacks, authentication hooks, and element system for serving files and custom UI components.

Element File Handling

When elements are sent, Chainlit copies server files into per-session folders, assigns UUID Chainlit keys, stores metadata in a database, and delivers files to clients via WebSocket retrieval requests.

Custom Element State

Custom JSX elements maintain mutable state persisted server-side. Client updates send full element payloads back for deserialization and storage. The talk walks through this update flow as part of tracing how AI framework weaknesses can lead toward exposure of cloud secrets.