Proof-of-concept of the danger of using neural networks for secure hashing (not that anyone does).
The idea to show that neural network hash functions expose critical information in their gradients came from seeing the following paper (https://www.researchgate.net/publication/310624366_Hash_function_generation_by_neural_network) that was cited in https://ai.google/research/pubs/pub46518.