There's No Secure Way to Use AI Agents Right Now

There’s no secure way to use AI Agents right now. Its important to understand this, specially for non technical people.

Even with all advancements like sandboxes, credential proxies and the models getting better, the fact that they don’t separate instructions from data and are non deterministic, make any agent prone to Prompt Injection Attacks. Specially for the Claw-like systems (OpenClaw, NanoClaw, etc) that is built to manage personal data like email and calendar.

Don’t be like the Meta Security Research that got all emails deleted by OpenClaw. As you give more of your personal data to those Agents, think on the worse scenarios, and take actions to mititage them. Maybe a daily backup of her inbox could help restore the previous state for example.