BlueHat IL 2026 - Oran Avraham, Ori Lahav - Exploiting Zero-Days via PostgreSQL Internals

In this presentation, we introduce the audience to the world of vulnerabilities in AI orchestration platforms. We researched some of the key players - Langflow, n8n, and Activepieces - uncovering multiple zero-day vulnerabilities across these ecosystems. We will show an RCE, a .tar symlink vulnerability in a chatbot turned into server takeover, and how knowing database internals and system design turns simple vulnerabilities into powerful attacks: A single mistake in JWT token validation forces an underlying ORM to drop WHERE conditions and return random files. To exploit this leak more reliably, we explore PostgreSQL internals, specifically sequential scans, and implement "heap-shaping" techniques within the PostgreSQL physical storage. Knowing system internals still matters in the new world of AI software.