Behaving appropriately under threat is key to survival. In my talk, I will provide a decision-theoretic view on this action selection problem and ask, what are computational algorithms and neural controllers that underlie this behavior. Non-human animal data tentatively suggest a specific architecture that relies on tailored algorithms for specific threat scenarios. To make this plausible in humans, I build on fear-conditioning paradigms, as well as on a translation of approach-avoidance conflict (AAC), a classical rodent anxiety model, to human computer games. I will analyze possible cognitive-computational algorithms for behavioral control and learning in these tasks, and their neural implementation.