/

CVE-2026-91932

Severity:

HIGH RISK

(8.5)

Input Validation Bypass (CWE-20) in Flowise MCP Server

CVE-2026-91932

Awaiting Analysis

Vulnerability type

Category:

Improper Input Validation

CWE:

CWE-20

Impact:

Remote Code Execution

Severity:

HIGH RISK

(8.5)

CVSS version:

3.1

Catch this CVE while code is being written

Detect vulnerable code directly in pull requests, so CVEs are fixed before merge.

Summary

Flowise contains a validation bypass in MCP server configuration that allows authenticated attackers to control an unvalidated working-directory parameter.

By combining a controlled cwd with clean filenames in the args array, an attacker may execute malicious code remotely under the privileges of the Flowise process.

Affected context

The vulnerability occurs in Flowise deployments that allow authenticated users to configure MCP servers and process working-directory or argument values. Risk is present when the cwd parameter is insufficiently validated and validation logic can be bypassed through otherwise acceptable-looking filenames in the args array.

Exploit preconditions

  • Attack surface: Authenticated MCP server configuration functionality
  • Required access: A valid account with permission to create or modify MCP server settings
  • Trigger point: Processing or launching an MCP server configuration
  • Why exploitation is possible: The working directory is not adequately constrained, while validation can be bypassed by supplying clean filenames through the argument configuration

Why this happens

The underlying bug is improper input validation, mapped to CWE-20. The application applies incomplete or inconsistent checks to related configuration fields, allowing attacker-controlled values that should be restricted by path, location, or execution policy.

Treating the cwd and args values independently creates a validation gap: individually acceptable values can be combined to produce an unsafe execution context. This allows configuration data to influence process execution without sufficiently enforcing trusted directory boundaries.

Potential impact

  • An attacker may execute arbitrary code on the Flowise server
  • The application process and its accessible data could be compromised
  • Credentials, workflow definitions, or integration secrets may be exposed
  • The host or surrounding services could be used for lateral movement
  • Workflows or stored data could be modified or poisoned

General mitigation guidance

Mitigation generally focuses on enforcing a single, consistent security policy for MCP server paths and arguments.

Common defensive patterns include:

  • Restricting working directories to explicitly approved locations
  • Canonicalizing and validating paths before any process is started
  • Treating cwd, executable paths, and arguments as a single validated configuration
  • Using allowlists for permitted MCP servers and launch parameters
  • Avoiding execution of user-controlled files or commands
  • Applying least-privilege permissions to the Flowise service account
  • Isolating MCP processes with suitable operating-system or container boundaries
  • Reducing the trust and mutability of configuration inputs

How teams detect issues like this

Detection can combine static analysis, review practices, and runtime policy enforcement:

  • Search for process-spawning code that uses user-controlled working directories or arguments
  • Flag path checks that validate fields independently rather than as an execution policy
  • Review canonicalization, normalization, and allowlist logic for bypass conditions
  • Monitor MCP configuration changes, process launches, and unexpected working directories
  • Enforce policies that restrict child-process creation and filesystem access
  • Use automated security review tools to identify unsanitized configuration values reaching process execution APIs

Sources

SourceLink
github.comgithub.com/FlowiseAI/Flowise/security/advisories/GHSA-x7x8-95gh-42xm
vulncheck.comvulncheck.com/advisories/flowise-before-3.1.4-remote-code-execution-via-cwd-parameter

Frequently asked

questions

It can be exploited remotely through Flowise's authenticated MCP configuration functionality when the attacker has the required permissions to create or modify an MCP server definition.

Yes. The described attack requires an authenticated account with access to the relevant MCP server configuration features. The practical risk also depends on the privileges assigned to that account.

Validate and canonicalize the working directory and all execution-related arguments together, restrict them to approved locations and values, and prevent untrusted configuration from controlling process execution.

Use strict allowlists, consistent validation at the final execution boundary, least-privilege service accounts, and monitoring for unexpected MCP configuration changes or child processes.

Find out what's already
exploitable in your codebase.