Although dogs exhibit remarkable capabilities for interacting with humans, the underlying cognitive mechanisms remain insufficiently understood. Here, we investigated canine perspective taking by challenging dogs to decide whether and where to steal food in the absence of a human who had prohibited them from doing so. The dogs could only infer the experimenter's presence through a sound (carrot chopping) they had perceived during a prior exploration phase, in which they also had the opportunity to observe from which locations in the room the human was visible. In the test, the majority of dogs preferred to steal from a plate that was not visible from the location where the human had chopped carrots before when they heard a playback of the chopping sound but not when they heard a control sound (street noise). These findings provide evidence that dogs anticipate the behavior of humans without relying on observable visual cues.