Critical Vulnerabilities Discovered in AI Orchestration Platforms

Research presented at DEFCON 34 revealed 14 critical and high severity vulnerabilities across seven AI orchestration platforms, including NocoBase, Flowise, Langflow, Dify, Activepieces, Kestra, and Apache Airflow. These platforms, integral for building AI agents and automating workflows, operate under the dangerous assumption that any user who can access a workflow is trusted to execute code on the host. Attack vectors include unauthenticated prompt injections leading to remote code execution (RCE), with some platforms allowing exploitation without user authentication. Notably, Flowise's vulnerability chain involves unauthenticated requests that can lead to RCE through pre-imported libraries. The findings highlight a significant security oversight in multi-tenant environments designed as single-user tools, where sensitive information can be compromised. The full technical whitepaper detailing these vulnerabilities has been made publicly available.

Key Points: • 14 critical vulnerabilities identified across seven AI orchestration platforms. • Vulnerabilities allow unauthenticated users to execute code remotely. • Research findings emphasize the need for improved security in multi-tenant environments.