We critically analyze an adversarial neural cryptography paper by Abadi that introduces a new cryptosystem, whose underlying principles are based on techniques of machine learning. While the approach has many positive features, our research indicates key shortcomings. We identify the main weak points: information leaking, adversarial attacks, and model robustness which are necessary for a secured cryptosystem. We suggest improvements geared toward countering these weaknesses and making the overall cryptographic architecture more secure as an outcome of in-depth analysis and testing. Our results open up other avenues of future development within adversarial neural cryptography systems that are even more resilient and also expand the frontiers of the security environment much better.