AI model hacked a custom 32-bit HellGates processor

Researchers tested the resilience of HellGates — a custom 32-bit processor with memory encryption support developed in VHDL language. During the experiment, the GPT-6 model was able to fully compromise the system in less than 30 minutes, demonstrating the vulnerability of the hardware architecture to automated analysis. In the course of the attack, the neural network used a combination of methods to bypass protection. The main vectors were the exploitation of weak cryptographic algorithm implementations and the application of side-channel analysis techniques. This allowed the model to successfully extract 1 GB of data from the processor's protected memory. For comparison, a human team participating in a similar challenge spent 4 hours solving the task and scored 14 points. The automated approach reduced the attack execution time by 8 times, highlighting the critical risks in designing hardware systems using non-standard encryption methods. The experiment's results sparked active discussion in the Hacker News community, where experts noted the need to revise security standards for gate-level design. This case confirms that modern LLMs are capable of effectively conducting reverse engineering of hardware, identifying logical errors in code at the gate level.