Executive brief
Flowise is an open-source visual AI workflow builder that allows users to create chatflows and agents. The CSVAgent component in Flowise accepts a user-supplied custom Pandas code parameter without sanitization, allowing an authenticated attacker to inject and execute arbitrary Python commands on the server. This can lead to complete server compromise and unauthorized access to sensitive data or system resources.
Technical details
The vulnerability is a code injection flaw in CSVAgent.ts where the customReadCSVFunc parameter is directly interpolated into Python code without sanitization and then executed via pyodide.runPythonAsync(). An authenticated attacker can craft a malicious payload (e.g., DataFrame({'foo': ['bar!']});import os;os.system('whoami')) within the customReadCSV parameter of a chatflow. The attack requires authentication to create a chatflow via POST /api/v1/chatflows, but authentication can be bypassed if default credentials are not set by providing the "x-request-from": "internal" header. Once the chatflow is created, sending a POST to /api/v1/prediction/[CHATFLOWID] executes the injected code on the server. Patches are available in Flowise 3.1.0 and flowise-components 3.1.0.
Affected products
- Flowise Flowise <= 3.0.13
- Flowise flowise-components <= 3.0.13
Timeline
- 2026-04-15: disclosed
- 2026-04-16: patched: Flowise 3.1.0 and flowise-components 3.1.0