DifyTap: Silently Wiretapping Multi-Tenant AI

Field CISO

Dify AI is a leading open-source LLMOps platform powering over one million applications across 60+ industries. With over 150,000 GitHub stars and 10 million Docker pulls, it has become critical AI infrastructure spanning self-hosted deployments, enterprise software, and a shared multi-tenant SaaS cloud. However, AI deployed in SaaS environments aggregate private chat logs, internal APIs, and uploaded documents;therefore, a single vulnerability on the server acts as a force multiplier, turning a minor vulnerability into a massive, cross-tenant exposure.

‍

In this talk, we present DifyTap: four vulnerabilities in Dify AI, including two critical severity flaws (CVSS 9.4 and 9.1) and two requiring no authentication. We will explain how analyzing the system architecture, including key differences between local, enterprise, and SaaS deployments, led us to uncover these vulnerable endpoints and code paths, and how these findings impact real-world production environments.

‍

We will walk through three core attack vectors: an unauthenticated path traversal in the plugin system that granted access to internal Plugin Daemon APIs (CVE-2026-41948), broken access controls in the file pipeline that exposed private enterprise documents and enabled unauthorized prompt extraction, and a tracing configuration flaw (CVE-2026-41947) that allowed an attacker to connect an external provider and silently wiretap live AI conversations across tenants.

‍

DifyTap serves as a case study for securing modern AI infrastructure. Beyond prompt engineering and probabilistic attacks, the greatest risk often lies in the traditional software layer managing the AI pipeline. Attendees will walk away with a fresh perspective on the LLMOps attack surface, along with practical methodologies for auditing multi-tenant AI control planes and identifying hidden architectural flaws before silent exploits occur.

‍