Securing AI-Augmented Liferay: Threat Modelling MCP, RAG Pipelines and Agentic Integrations
Presenters:
- Hong Sheng, Lead Solution Architect, Liferay
- Mohit Rajput, Solution Architect, Liferay
Liferay DXP ships with a robust security foundation Content Security Policy, OAuth2, role-based permissions, and identity management built in. But every AI integration added on top of that foundation opens new attack surfaces that none of those controls were designed to address.This session covers few of them:1. Prompt injection via MCP — portal content that hijacks agent instructions, bypassing CSP and permissions entirely because the attack travels through LLM context, not the browser2. RAG pipeline data leakage — documents in the DXP library that cause an LLM to exfiltrate data from unrelated retrieved chunks, circumventing folder-level role restrictions3. Vibe coding vulnerabilities — SQL injection, hardcoded credentials, and missing PermissionChecker calls introduced when AI-generated Liferay code ships without critical review4. Broken object-level authorisation — agent OAuth2 tokens scoped too broadly, reaching content and users across sites they were never intended to access5. Event pipeline poisoning — crafted web-hook writing XSS or fabricated signals into DXP Objects, evading output encoding at the sourceFor each, the session covers what layered mitigation does — from CSP hardening and OAuth2 scope restriction to MCP tool auditing, context sanitisation, and event schema enforcement.Attendees leave with a prioritised checklist of controls applicable to any Liferay AI integration.