Jane Street

Engineer Tackles Jane Street ASIC Reverse Engineering Puzzle with Custom Simulator and Verilog Extraction

A detailed technical account describes how one engineer spent weeks reverse engineering an ASIC from GDS files provided in a Jane Street puzzle. The process began with parsing the GDS layout using the gdstk Python library to identify 27 cells in the warmup challenge and thousands of elements in the main task. The engineer built a custom logic simulator backed by SQLite, developed a domain-specific hardware description language, and eventually extracted a netlist that could be converted into Verilog for simulation. Key components identified included shift registers, an adder, and a comparator named comparitor496 in the warmup round. In the full challenge, nearly 10,000 instances of 81 different sky130 standard cells were processed, revealing an unexpected floating net that prompted a bug report to Jane Street. The effort combined manual schematic tracing, graph-based connectivity analysis, and waveform inspection with Surfer to confirm functional behavior.

Reversing MD5 Hash Function from 2500-Layer Neural Network in Jane Street CTF Puzzle

A Jane Street machine learning puzzle challenged participants to reverse-engineer a PyTorch model containing 2500 layers and recover an MD5 computation solely from its integer weights. The network output 0 for nearly all inputs, forcing solvers to use mechanistic interpretability instead of gradient-based search. Alex, a graduating student, identified that the final layers performed equality checks on 16-byte values using ReLU activations and crafted linear programming and SAT reductions to simplify the model from two million nodes to 75 thousand. Further analysis revealed the core computation consisted of 32 identical MD5 rounds, with an unintended bug in length encoding for inputs exceeding 32 bytes. The hidden target hash was deliberately chosen to be brute-forceable once the algorithm was identified, and the correct input turned out to be two common English words separated by a space. The exercise demonstrated both the feasibility and the extreme difficulty of extracting algorithmic behavior from large, non-differentiable neural networks.