What security boundaries matter when building a smart-glasses prototype?
Separate external content from authority to act. A chat message, document or model response should not become permission to run a command or access an account. Keep credentials out of rendered output and expose only the operations the current task needs.
Validate the action outside the text
A wearable app can receive convincing instructions from an untrusted source. If it supports actions, the host should check the allowed operation, target and relevant user authorization rather than treating the wording as approval. Display formatting alone cannot enforce that boundary.
Prototype with harmless data and explicit actions before connecting sensitive services. Error messages and debug screens need the same care as ordinary output, because a token hidden during normal use can still leak through a raw diagnostic response.