The new mode runs WhatsApp’s Meta AI in the company’s Private Generation enclave, where conversations are deleted by default and no server-side logs are kept.
Meta has launched a Secret Chat mode for Meta AI on WhatsApp and Meta AI, an effort to overcome the awkward fact that, like every other major AI chatbot, its assistant has so far been able to read users’ conversations with it.
new mode, the company announced Tuesdayprocesses user messages in what Meta describes as a secure environment that even Meta can’t see, with chats deleted by default after the session ends.
The technical core is WhatsApp’s Private Generation system, an architecture the company published in April 2025 to allow AI features to run on encrypted data within Trusted Execution Environments on Meta servers.
Inside that enclave, the model can read and respond to the request, but its contents are not accessible to Meta’s engineers, its input systems, or any commercial pipelines.
Other apps offer incognito modes for conversations with AI, but the announcement highlights the Meta frame: “they can still see the questions coming in and the answers coming out.”
The startup directly addresses the privacy challenge across categories. AI chatbots have become a standard tool for questions users would once ask a doctor, lawyer or partner.
OpenAI, Google, and Anthropic each store chat histories by default with various user controls. Apple Intelligence routes some requests through Apple’s Private Cloud Computing, an enclave architecture that is the clearest analogue to what Meta sends within WhatsApp.
Two product details come from the design. First, conversations are generally not stored server-side; users can no longer raise their Secret Chat history because there is nothing to raise.
Second, the disappear-by-default behavior means that even a compromised device is less likely to leak, because conversation remnants are cleaned up between sessions.
Meta published technical document description of the cryptographic architecture for external viewing.
A second feature is on the way. Side chat with Meta AI, also protected by Private Generation, will allow users to receive AI assistance within an existing WhatsApp conversation, the assistant is aware of the context of the chat, but its responses remain invisible to other participants.
Meta said Sidechat will come to WhatsApp “in the coming months” without a more specific date.
The commercial logic of the startup is simple. WhatsApp has been built around end-to-end encryption as a selling point for a decade, and Meta’s bid for AI on the platform had to find a way out of the central tension that a conversational AI assistant needs to read your messages to be useful.
Private Processing is the company’s attempt to square that circle. The Incognito Chat product is the first time that architecture has been put behind a user-facing feature at this scale.
Whether the implementation is under control is a separate question. Trusted Execution Environment-based AI systems have been scrutinized and criticized across the industry, with researchers periodically demonstrating side-channel attacks against similar architectures from Apple, Google, and hyperscalers.
Meta has invited external review of its Private Processing design, and the new white paper expands on that, but the model’s resistance to challenge, in particular, has yet to be tested in court.
Secret chat with Meta AI is rolling out this week on WhatsApp and the Meta AI app, with more widespread use in the coming months.
The release comes at the end of a difficult two weeks for Meta on the privacy front, the company’s new US employees mouse tracking software on Monday, and the company is coming off a week of layoffs of nearly 8,000 workers.
Within the meta, consumer-facing privacy moves like this are likely to trump internal control optics.
