What is Secret Alignment? Introduced by me and Prof. Jung-Eun Kim in early 2025, Secret Alignment describes a class of hidden, condition-dependent behaviors in aligned models — the model's response is systematically different under a specific trigger, context, or credential, while remaining indistinguishable from a normally aligned model to any observer without that trigger.

Because the same mechanism can serve legitimate control or hostile subversion, Secret Alignment is fundamentally dual-use:

Whether Secret Alignment serves defense or attack depends on who holds the trigger — which is why my work treats it as a first-class object of study and evaluation.