Attack Chain

1. An attacker hosts a malicious web page designed to interact with the Network-AI MCP server.
2. A user with a default-configured Network-AI MCP server running locally visits the malicious web page.
3. The malicious web page sends an HTTP OPTIONS request to the /mcp endpoint to check CORS preflight. The server responds with Access-Control-Allow-Origin: *.
4. The malicious web page sends an HTTP POST request to the /mcp endpoint with a JSON-RPC payload targeting a MCP tool (e.g., config_set). No Authorization header is included.
5. The server’s _isAuthorized function evaluates to true because the secret is empty.
6. The server’s _handleRequest function sets Access-Control-Allow-Origin: * on the response.
7. The server’s _bridge.handleRPC function executes the requested MCP tool (e.g., config_set to modify configuration).
8. The malicious web page receives the response and can read the result due to the permissive CORS setting, confirming successful execution of the MCP tool.

Impact

Any web page visited by a user who has the Network-AI MCP server running locally on the default port (3001) with no configured secret can silently invoke all 22 MCP tools without credentials. This allows for arbitrary orchestrator configuration mutation (config_set), spawning arbitrary agents (agent_spawn), corrupting shared agent state (blackboard_write / blackboard_delete), and tampering with token management (token_create / token_revoke). The integrity impact is high because core orchestrator state can be overwritten.

Recommendation

• Apply the vendor’s suggested remediation by requiring a non-empty secret at startup (see remediation #1 in the overview) to prevent unauthorized access.
• Implement the vendor’s suggested fix by restricting CORS to localhost origins only (see remediation #2 in the overview) to prevent cross-origin requests.
• Deploy the Sigma rule “Detect Network-AI MCP Tool Invocation Without Authorization” to identify attempts to exploit this vulnerability by monitoring POST requests to the /mcp endpoint without an Authorization header.
• Upgrade to a patched version of Network-AI that addresses CVE-2026-46701.