Anonymizing proxy

A proxy sits between you and a service. Your request goes to the proxy, the proxy calls the service with its own address and credentials, and the answer travels back the same way. Corporate gateways, privacy relays and forwarding services all follow this pattern. The service sees the proxy, not you, much like a mail-forwarding address: the sender learns where the forwarder lives, not where you do, yet can still read the letter inside. A large lab can thus receive a prompt with no name, email, phone number or network address attached.

The limit is the content. A model must read a message to answer it, so the text itself reaches the service in readable form. If you write your name, your employer or your health situation in the message, the service reads that, whatever the proxy hides around it. Even a harmless-looking detail, a rare job title plus a small town, may single someone out, which is why identity and content have to be considered separately. Some proxies add a masking step: recognizable personal data is swapped for placeholders before sending, and the real values return in the answer. Masking catches patterns, but not what a sentence can reveal about a job, a town or a diagnosis. Journalists, researchers and security teams use relays like this to read sensitive sources or test services without leaving a trail that points back to them.

A proxy relocates trust instead of removing it: you rely on its promises about what it logs and forwards. It is different from a VPN, which hides your network address but leaves you signed in with your own account, and different from encryption, which protects data in transit but does not stop the destination from reading it.

This is what the Anonymized label means on NeverNo. For Claude, GPT, Gemini and Grok, our server calls the provider with its own key. The maker receives no email, account id or IP address of yours; it reads the text of the conversation without knowing who you are, applies its own rules and may keep the request for a while under its own policy. Before a message leaves your browser for one of these models, a closed list of personal data (email addresses, phone numbers, bank and card numbers, IP addresses, postal addresses written in full, the first name you gave and the names you add) is replaced by placeholders like [EMAIL_1], and the real values are put back in the reply. Any other name, a place or an employer passes as written. A switch in Settings, on by default, controls this. The transparency page publishes measured detection rates, so you can judge what masking catches; group chats, missions and the public API do not go through it.

GPT-6.1 Sol, OpenAI's balanced model, shows the idea at work. Its context holds 200,000 tokens, it reasons on every answer and reads images. OpenAI never receives your name, account or IP address through NeverNo, yet reads what you write and applies its own rules. Ask Sol to draft a contract clause or summarize a report and it works as it would anywhere else; only the sender's identity is held back. For a sensitive conversation, the transparency page advises staying on a private model, the open models served by providers that keep nothing.