Show HN: Decypher-env, an RL Env for breaking AES encryption (github.com)

A new reinforcement-learning (RL) environment called Decypher-env has been introduced, designed to tackle the complex task of breaking AES-XTS encryption by generating C++ solvers. This innovative setup presents a reduced-round AES inversion challenge, allowing developers to train models that learn to emit code to solve increasingly difficult instances of the encryption task. The environment operates with a clear reward structure, scoring solvers based on their ability to converge on solutions while also incorporating systems for compiling and executing their outputs. The significance of Decypher-env lies in its potential to advance cryptanalysis and machine learning techniques within the AI/ML community. By providing a structured curriculum and adaptable interface for major RL frameworks, the environment fosters experimentation and benchmarking in a challenging domain. The architecture caters to RL developers by allowing them to track progress through qualitative signals and standardized reward metrics, paving the way for improving solver performance. The tool also emphasizes the complexity of full-round AES-XTS inversion, setting realistic expectations while creating opportunities for breakthroughs in understanding cryptographic resilience through machine learning methods.