WhatsApp says Incognito Chat runs Meta AI inference inside Trusted Execution Environments so neither Meta’s operators nor others can access conversation contents; sessions are ephemeral and text-only for now.
WhatsApp on Wednesday introduced Incognito Chat, an AI chat mode that runs all Meta AI inference inside a Trusted Execution Environment (TEE) as part of its Private Processing architecture. The company says the setup prevents operators — including Meta itself — from accessing the content of those exchanges, giving users a way to query the assistant without exposing questions or answers to the service provider. This matters for users who want AI help while keeping conversation contents private.
Incognito Chat extends Private Processing, a program WhatsApp launched about a year ago that already supports message summarization and composition. Conversations in Incognito Chat are ephemeral by default and disappear when a session ends; the feature is text-only for now. Meta says it will be able to detect only that an account used the capability, not the specific prompts or responses.
Meta frames the TEE approach as different from typical “incognito” modes offered by other AI services, which may separate identity from logs but still retain server — side records. Will Cathcart, head of WhatsApp, described the trade — off as running large models in a secure data-center environment rather than fully on-device, calling it “like we’re running a giant phone for AI and we don’t have the passcode.
The company is inviting third — party audits and external vulnerability reports for Private Processing; it says Incognito Chat will be subject to expert oversight so outside parties can verify the shipped code and attest to its protections. Meta CEO Mark Zuckerberg posted that inference inside the TEE ensures conversations aren’t accessible to Meta and reiterated that chats on the phone disappear after a session.
Alongside Incognito Chat, Meta announced Side Chat with Meta AI, a feature that lets users ask the assistant about an existing text conversation — for example, requesting restaurant suggestions or a summary without manually copying messages into a separate chatbot. Meta also plans to add Incognito Chat access within the Meta AI app, to offer an opt-out default for anonymized web search the model may use, and to work on image and voice support.
For builders and security teams the rollout highlights concrete engineering requirements: integrating inference inside TEEs, optimizing routing and latency to keep cloud — based private inference responsive, and preparing for external code review or attestations. Meta cautions that, as with any cloud service, users and developers still place trust in the provider’s underlying infrastructure even when technical controls limit direct access to message contents.
Sources
Replies (0)
No replies in this topic yet.
