A hidden malicious behavior inserted into a model through low-rank adaptation weights that triggers on specific inputs.