Secure Messaging and AI Remain In Conflict Despite the Promise of TEEs
4 hours ago
- End-to-end encryption secures messages in transit, but once on a device, AI features may process data in ways that compromise privacy.
- On-device AI is less concerning, but cloud-based AI often requires sending data to servers, which TEEs aim to protect.
- TEEs (trusted execution environments) are hardened server sections that run code privately, but they rely on engineering, not math, making them less secure than encryption.
- TEEs have been repeatedly hacked via side-channel attacks, unlike encryption, which is mathematically robust and unbroken to date.
- Companies like Apple, Google, and WhatsApp use TEEs for private AI processing, but this differs from actual encryption.
- The flaw: TEEs can be physically accessed, and once the key is on the device, attacks become possible, unlike encryption where keys are never on the server.
- Homomorphic encryption could resolve privacy but is too slow for LLMs, pushing companies to use TEEs as a compromise.
- Users should avoid automatic data sending to TEEs; manual control allows them to assess risk for sensitive messages.
- Automatic exfiltration to TEEs downgrades end-to-end encryption to a weaker security model.
- The industry trend risks eroding privacy guarantees for secure messaging, confusing users about what data is actually protected.