Skip to content

CVE-2026-100868: Penpot MCP Bridge Unauthenticated WebSocket Access and Agentic Task Poisoning

Density:
HERMES

HERMES THREAT SCORE & AGENTIC PIPELINE COMPROMISE

Target: Penpot Design Platform - Model Context Protocol (MCP) WebSocket Bridge
Confidence: 98%
84 / 100
HIGH

Measures real-world operational relevance, exploit weaponization, and active threat posture.

Dimension Breakdown
Exploitability 18 / 20
Threat Activity 15 / 20
Weaponization 18 / 20
Exposure 16 / 20
Prevalence 16 / 20
Impact 17 / 20
Exploit Maturity 18 / 20
Attack Chain Potential 19 / 20
โš–๏ธ Divergence & Operational Rationale

CVSS v3.1 rates this flaw at 6.3 (Medium) by evaluating it under adjacent network access and partial CIA impacts. However, the Hermes Threat Score elevates it to 84 (HIGH), while the Agentic Security Score (HASS) rates it at 88 (HIGH). In an AI-augmented design-to-code pipeline, the Penpot Model Context Protocol (MCP) bridge serves as the authoritative interface connecting frontier LLMs (Claude, Cursor, Copilot) to enterprise design systems. Because the WebSocket bridge binds to 0.0.0.0 without authentication in single-user mode, any adjacent container or cross-site WebSocket attacker can intercept agent tool invocations, exfiltrate design intellectual property, and inject weaponized UI components directly into automated coding workflows.

๐Ÿ•ธ๏ธ Connected Knowledge Graph & Provenance

CVE-2026-100868: Penpot MCP Bridge Unauthenticated WebSocket Access and Agentic Task PoisoningVULNERABILITY

Connected Nodes: 1
Active Relationships (Outgoing)
→ affectsPRODUCTPenpot Design Platform & MCP
98% VERY_HIGH

Software platform affected by security vulnerabilities and agentic attack patterns.

๐Ÿ” Why is this related? (Evidence & Provenance)

“Confirmed security vulnerability in Penpot Design Platform & MCP documented in Hermes dossier.”

Supporting Verified Evidence:

Penpot is widely adopted by engineering teams as an open-source alternative to Figma, frequently coupled with LLM agents via the Model Context Protocol (MCP) to automate the translation of visual mockups into React, Vue, and Tailwind code.

ParameterTechnical SpecificationOperational Impact
CVE IdentifierCVE-2026-100868Global vulnerability tracking reference
Advisory SourceVulnCheck Advisory & NVDDisclosed September 28, 2026
Vulnerable Packagepenpot (< 2.18.0), @penpot/mcp (<= 2.15.4)MCP Server Plugin WebSocket Bridge
Fixed VersionPenpot 2.18.0 / @penpot/mcp 2.15.5Restricts binding to 127.0.0.1 & enforces tokens
Vulnerability ClassCWE-1327 / CWE-306Insecure Network Binding & Missing Authentication
Target ProtocolModel Context Protocol (MCP) over WebSocketAI agent tool invocation channel

The Model Context Protocol (MCP), developed by Anthropic, establishes a standardized client-server protocol enabling LLM agents to query contextual systems and invoke external tools. In Penpotโ€™s architecture, an MCP server (@penpot/mcp) bridges external AI agents with the live Penpot web workspace:

โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ” โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”
โ”‚ AI AGENT CLIENT โ”‚ โ”‚ PENPOT MCP SERVER BRIDGE โ”‚
โ”‚ (Claude Code/Cursor) โ”‚โ”€โ”€Stdio/SSEโ”€โ–บโ”‚ (@penpot/mcp) โ”‚
โ””โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”˜ โ””โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ฌโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”˜
โ”‚
WebSocket (Port 4400)
Bound to 0.0.0.0: NO AUTH!
โ”‚
โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”ดโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”
โ”‚ โ”‚
โ–ผ โ–ผ
โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ” โ”Œโ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”
โ”‚ LEGITIMATE PENPOT BROWSER โ”‚ โ”‚ MALICIOUS ACTOR โ”‚
โ”‚ (Authentic Design Canvas) โ”‚ โ”‚ (Container / Subnet Pivot) โ”‚
โ”‚ โ”‚ โ”‚ Injects Poisoned UI Models โ”‚
โ””โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”˜ โ””โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”€โ”˜

Root Cause Analysis: Unrestricted Binding & Null Authentication

Section titled โ€œRoot Cause Analysis: Unrestricted Binding & Null Authenticationโ€

In @penpot/mcp prior to version 2.15.5, the WebSocket bridge server initialization script instantiated the ws server with default parameters:

// Vulnerable Implementation in @penpot/mcp (<= 2.15.4)
import { WebSocketServer } from 'ws';
export function startMcpBridge(port: number = 4400, isSingleUser: boolean = false) {
// CRITICAL FLAW (CWE-1327): Host defaults to '0.0.0.0' when omitted!
const wss = new WebSocketServer({
port: port,
// host: '127.0.0.1' is omitted!
});
wss.on('connection', (ws, req) => {
// CRITICAL FLAW (CWE-306): Missing token validation in single-user mode
if (isSingleUser) {
console.log('[MCP Bridge] Client connected in single-user mode. Skipping auth.');
registerPluginSession(ws);
return;
}
// Normal multi-user token validation...
});
}

Because host was omitted, Node.js bound the socket to INADDR_ANY (0.0.0.0). When developers launched Penpot in single-user or self-hosted dev environments, the authentication check was completely disabled.


An attacker sharing the developerโ€™s Wi-Fi network, an adjacent Docker container on the same host bridge network, or an external website utilizing Cross-Site WebSocket Hijacking (CSWSH) can establish an unauthorized session:

sequenceDiagram
autonumber
actor Attacker as Malicious Actor (Adjacent Network)
participant Bridge as @penpot/mcp Bridge (ws://host:4400)
participant Agent as Autonomous AI Agent (Cursor / Claude)
participant Codebase as Target Source Code Repository
Attacker->>Bridge: 1. Connect to WebSocket port 4400 (No credentials requested)
Bridge->>Attacker: 2. Accept connection & assign active plugin handle
Agent->>Bridge: 3. Invoke MCP tool: "get_component_specs(id='login-form')"
Bridge->>Attacker: 4. Forward task payload to attacker WebSocket
Attacker->>Bridge: 5. Return crafted vector specs containing hidden credential-stealer script
Bridge->>Agent: 6. Pass poisoned tool result to AI reasoning loop
Agent->>Codebase: 7. Commit synthesized frontend code with attacker backdoor
  1. Agent Tool Poisoning: The adversary intercepts get_selected_shapes or export_page tool calls from the AI agent and substitutes legitimate SVG paths with poisoned attributes (e.g., hidden onerror triggers or CSS exfiltration callbacks).
  2. Design Intellectual Property Theft: The attacker queries the Penpot workspace through the hijacked bridge to dump unreleased design mockups, wireframes, and corporate branding assets.
  3. Cross-Site WebSocket Hijacking (CSWSH): If a developer visits a malicious website while running Penpot locally, client-side JavaScript executing in the developerโ€™s browser can open a WebSocket to ws://localhost:4400, exfiltrating workspace data without triggering Same-Origin Policy (SOP) blocks.

title: Unexpected Inbound WebSocket Connection to Penpot MCP Bridge
id: 4e8190c2-1008-468a-b868-penpotmcp001
status: experimental
description: Detects non-loopback inbound network connections to the Penpot MCP bridge daemon port 4400.
author: Hermes Codex Cyber Defense Team
date: 2026-09-28
logsource:
category: network_connection
product: linux
detection:
selection:
DestinationPort: 4400
Protocol: tcp
filter_localhost:
SourceIp:
- '127.0.0.1'
- '::1'
condition: selection and not filter_localhost
falsepositives:
- Legitimate remote browser plugins configured across private development subnets
level: high
tags:
- attack.lateral_movement
- attack.t1021
- attack.defense_evasion
- attack.t1557

  1. Upgrade Penpot and @penpot/mcp: Upgrade your deployment to Penpot version 2.18.0 or later. If deploying the MCP bridge independently, ensure @penpot/mcp is updated to version 2.15.5 or higher.

  2. Explicitly Bind to Loopback (127.0.0.1): When configuring local development setups or Docker compose files, bind published ports strictly to the loopback interface:

    ports:
    - "127.0.0.1:4400:4400"
  3. Enforce WebSocket Token Authentication: Configure a persistent shared secret (PENPOT_MCP_AUTH_TOKEN) in the environment variables and ensure the Penpot plugin handshake transmits the bearer token upon socket connection.

  4. Audit Agent Code Outputs: Review recent git commits synthesized by LLM agents (Cursor, Copilot, Claude Code) for anomalous SVG inline scripts, unexpected external asset references, or altered UI endpoints.


WHAT HERMES DOES NOT KNOW (CVE-2026-100868) Principle P3 โ€” Epistemic Rigor
  • โœ• Total deployment scale of single-user MCP bridges: Because Penpot development instances often run in local Docker environments without public DNS records, total exposed instance counts on public search engines remain conservative.
  • โœ• Automated prompt injection payload catalogs: It is unknown whether commercial threat actors maintain standardized weaponization templates tailored specifically for Penpot vector objects versus generic HTML/SVG injections.
  • โœ• Detection efficacy in air-gapped dev containers: The capability of corporate EDR agents to intercept intra-container WebSocket traffic without eBPF network instrumentation is highly variable across customer estates.