08/10/2026
If an AI reads your business email, this is for you. ⚠️
1. The most dangerous email is not written for a human. Plugging AI into the inbox is powerful, and it opens a door most businesses have not heard of yet.
2. How it works: AI assistants follow written instructions, and email lets strangers send them. Attackers hide instructions for the AI inside an ordinary-looking message. The reader sees a normal email; the assistant sees orders. That is prompt injection.
3. Proof it is real: EchoLeak. One crafted email could make Microsoft 365 Copilot hand over internal files, with nobody clicking anything. Severity 9.3 out of 10, patched before known harm, but the method keeps resurfacing in new forms.
4. The guardrails: limit what the assistant can read, keep a human approving anything it sends or deletes, patch on discipline, allow no unvetted plugins, and make sure your AI policy covers assistants.
5. The point: AI in the inbox is a deployment decision, not a toggle. That deliberate deployment is exactly what we do for clients.
Worth checking today: has an AI quietly arrived in your inbox by default?