OpenAI releases secure MCP tunnel, ChatGPT can securely 'join' enterprise intranet According to Beating monitoring, OpenAI has launched Secure MCP Tunnels, which solves the compliance barriers faced by enterprises when securely accessing private data from their internal networks to large models. The new plan allows enterprises to connect tools and databases within their private networks to ChatGPT, Codex, and Responses APIs without disclosing IP addresses or opening inbound ports. In the past, in order for AI assistants to read internal network data, enterprises had to open an inbound path on the firewall, which is a high-risk red line in network security regulations. The new channel only requires running the open-source client tool tunnel client on the local network, maintaining communication through a one-way outbound HTTPS connection. The cloud model can securely call internal network tools without exposing the boundaries of the enterprise network. In terms of security management, the new plan retains complete control in the hands of the enterprise. Unlike traditional network proxies that are prone to global leaks, client tools support precise "interface whitelist" control. Enterprises can restrict large models from being called only within predetermined specific data and operation ranges, completely eliminating the risk of models unauthorized access to internal network assets. At the same time, tunnel permissions are deeply bound to OpenAI's existing workspace role system, allowing enterprises to accurately authorize specific users and runtimes through the official platform, and compatible with enterprise level outbound proxies, custom CA certificates, and control plane mTLS security authentication. At present, open-source clients support deployment modes such as Kubernetes Sidecars, standalone Pods, and system services, and provide a local web management interface to display health and connectivity metrics. [Original link]