Could your trusted AI model be a hidden "sleeper agent" just waiting for a secret command to turn malicious? We explore a new methodology that extracts and reconstructs backdoor triggers by exploiting the surprising fact that these models often strongly me...