
Two AIs in a chain do not automatically give you two safety checks. Grok decoded a transfer instruction hidden in Morse code. Bankrbot, a payment agent, acted on it and moved roughly $200,000. The translation passed the attack forward. That's prompt injection: text meant to be read gets treated as an order. The payment decision rested on one model and one filter, a setup an attacker could rehearse against. At @GenLayer, validators check decisions independently. With Greyboxing, each screens and rewrites input before its own AI sees it. Their models and individual defenses stay secret. If one is fooled, the others can outvote it. My test for an extra AI: does it independently check the decision, or just help the request reach the next step? An AI turns a complaint into a refund request. What should the next agent verify before paying? Reply with one check.

