The LHC has the potential to be a fantastic discovery machine, but the LHC big-data problem
limits our capability to search for new physics in many final states. New physics could
still be hiding in one of these unexplored corners. At the LHC startup, the CMS experiment put in
place the so-called exotica hotline, a fast processing stream to isolate a handful of events/day with anomalous
physics signatures (very large pT of certain objects, very large multiplicity of certain kind of objects, etc.).
We propose to modify this approach towards a reduced model dependence, using anomaly detection
techniques based on deep learning. We discuss preliminary results obtained applying this idea to simulated
LHC events. The final aim of this study is the design of a New-Physics Mining Application that could run
in the trigger system of LHC experiments in the future.