Secure Messaging and AI Remain in Conflict Despite the Promise of TEEs

Summary

Secure messaging platforms assume that message contents remain private to the people in the conversation, but AI features can weaken that boundary when they send data to cloud servers. On-device AI is less concerning, yet larger models and heavier workloads may require server-side processing. Companies including Apple, Google, and WhatsApp use trusted execution environments (TEEs) to process requests inside protected server components and use attestation to show which code is running. The article argues that this is privacy-preserving, but not equivalent to encryption or local computation. TEEs depend on engineering protections, contain keys on the same physical machine, and have repeatedly faced side-channel and other attacks; patches can reduce risk but cannot guarantee perfection. End-to-end encryption keeps the relevant keys off the service’s server, while fully private inference through homomorphic encryption remains too expensive for practical LLM workloads. Sending a conversation to a TEE for summarization, review, or storage therefore creates a meaningful downgrade from end-to-end encryption. The article recommends that devices never send such data automatically. Developers should require an explicit user choice, and users should disable automatic transfers and assess whether the data is worth the added risk.