Deficient safety learning has been implicated in the pathogenesis of anxiety disorders. Despite increased translational interest, there has been limited research on the basis of safety learning in humans. Here, we examined safety learning in seventy-three healthy participants via a modified Pavlovian conditioned inhibition paradigm, featuring a conditioned threat stimulus that was reinforced alone (A+), but not when combined with a second stimulus (the conditioned inhibitor, AX-). During a test phase, X and a control safety cue (C) were combined with a second threat stimulus to assess their inhibition of threat responses, measured via skin conductance (SCRs) and US-expectancy ratings. Both stimuli exhibited conditioned inhibition, but X suppressed ratings by a greater magnitude than C. Trait anxiety also predicted increased US-expectancy ratings of X. These findings suggest that a Pavlovian inhibitor accrues greater safety value than a merely unreinforced safety signal. Conditioned inhibition paradigms may have utility in the ongoing study of safety learning and its relevance to anxious psychopathology. •Safety learning was compared between a conditioned inhibitor and a conventional CS-.•Both inhibited threat responses at test, but the inhibitor showed greater suppression of US-expectancy ratings.•The inhibitor was easily re-learned as threatening when subsequently reinforced, contrary to conventional theory.•US-expectancy to the conditioned inhibitor was associated with higher trait anxiety.