Operational Security in CRDT-Based Collaborative Editing Systems

Conflict-free Replicated Data Types (CRDTs) enable real-time collaborative editing by allowing distributed participants to modify local copies of shared documents while ensuring eventual convergence. Although this architecture provides availability and synchronization benefits, it introduces significant operational security challenges when participants cannot be fully trusted. This presentation examines security limitations in CRDT-based collaborative editing systems, focusing on Yjs, Automerge, Loro, and associated synchronization infrastructures. Particular attention is given to the persistence of deleted sensitive information, document-level authorization constraints, client identity impersonation, synchronization failures caused by rejected operations, the potential loss of offline contributions, and limitations of participant revocation. The analysis incorporates technical demonstrations and reproducible experiments involving Yjs, Hocuspocus, y-websocket, and Durable Streams. These examples illustrate how malicious participants may manipulate operation identifiers, impersonate legitimate collaborators, and compromise the integrity of shared documents. A central contribution is the examination of artificial intelligence agents as active CRDT participants. The presentation demonstrates how AI-assisted editing can amplify existing authorization weaknesses through confused-deputy scenarios, excessive agency, and exposure to untrusted document content. It further discusses the implications of the Agents Rule of Two for collaborative AI systems. Practical security recommendations include separating untrusted content from editing privileges, introducing human review before applying AI-generated modifications, restricting agent capabilities through dedicated tools, assigning distinct agent identities, and implementing cryptographically verifiable operations. Fine-grained access control at the document-section level is also considered as a direction for improving collaborative editing security. The presentation highlights the need to integrate access control, identity verification, cryptographic integrity, and human oversight into the architecture of secure CRDT-based collaborative systems.