HellGates, custom CPU gate-level challenge

summary

The author created a highly complex, obfuscated 32-bit virtual CPU designed to be unsolvable by humans and AI, but it was eventually cracked by GPT-6 in under 30 minutes. The success of the AI was not due to a complete reverse-engineering of the logic gates, but rather the exploitation of a weak side-channel in the CPU's state encryption. This demonstrates that even extreme obfuscation can be bypassed if a single cryptographic weakness exists.

key points

1. The virtual CPU featured extreme protections, including VHDL-synthesized gate-level netlists, TEA-based memory encryption with dynamic keys, and host-side shellcode obfuscation using ChaCha20 and recursive exception handlers.
2. GPT-6 bypassed the layers of obfuscation by performing differential analysis on a lazily implemented state encryption loop, allowing it to decrypt memory and replay runs without fully reversing the hardware logic.
3. The author emphasizes that while the AI solved the puzzle quickly, it did so by finding the path of least resistance rather than solving the computationally infeasible task of analyzing millions of logic gates.

who should read it

Reverse engineers, security researchers, and AI enthusiasts interested in the intersection of hardware obfuscation and LLM capabilities.

discussion warm