Naze Pixel
A small C# sample application used for reverse engineering testing. It comprises roughly 105 lines of code and converts 24-bit BMP bitmaps into ASCII art through grayscale conversion, scaling and Floyd-Steinberg dithering. It contains an embedded upload key.
Timeline (last 90 days)
The test target Naze Pixel is about 105 lines of C# code, whose function is to convert 24-bit BMP into dithered ASCII art, and it has a built-in upload key.
In the ordinary .NET build test, REA's .NET reader returned 114 KB of JSON in about 1.5 seconds, including every method, signature, bytecode summary, and all string literals (including the key), and the AI rebuilt the entire application after 85 seconds.
In test one, the SHA-256 hash of the AI-rebuilt Naze Pixel code compiled was identical to that of the original program, proving they are byte-for-byte identical.
In the obfuscated build test, the open-source obfuscator Obfuscar was used to rename classes and methods to invisible Unicode characters and encrypt strings, but the embedded upload key still existed in plaintext in the obfuscated .dll.
In the obfuscated build test, the AI rebuilt the entire algorithm in about three minutes based only on REA output and cracked the XOR encryption (b ^ i ^ 0xAA), but it could not actually read the strings and could only guess by length.
In the obfuscated build test, the control group agent not using REA completed decompilation within 56 seconds, found the encryption logic, and decrypted all strings (including the key) with a few lines of Python.
In the Native AOT test, REA loaded the 1.1 MB pure x64 executable into Ghidra, Ghidra found 3627 functions (almost all of them .NET runtime), and the AI restored the entire application after about 18 minutes, including the same weights, dithering, and key.
In the Native AOT test, the control group agent not using Ghidra wrote its own disassembly script in Python in under 4 minutes, recovered the same constants, key, and real class names, and ran the unknown exe to observe its behavior.
All Facts (8)
In the Native AOT test, the control group agent not using Ghidra wrote its own disassembly script in Python in under 4 minutes, recovered the same constants, key, and real class names, and ran the unknown exe to observe its behavior.
50%UnverifiedIn the ordinary .NET build test, REA's .NET reader returned 114 KB of JSON in about 1.5 seconds, including every method, signature, bytecode summary, and all string literals (including the key), and the AI rebuilt the entire application after 85 seconds.
50%UnverifiedThe test target Naze Pixel is about 105 lines of C# code, whose function is to convert 24-bit BMP into dithered ASCII art, and it has a built-in upload key.
50%UnverifiedIn the obfuscated build test, the open-source obfuscator Obfuscar was used to rename classes and methods to invisible Unicode characters and encrypt strings, but the embedded upload key still existed in plaintext in the obfuscated .dll.
50%UnverifiedIn the obfuscated build test, the AI rebuilt the entire algorithm in about three minutes based only on REA output and cracked the XOR encryption (b ^ i ^ 0xAA), but it could not actually read the strings and could only guess by length.
50%UnverifiedIn the obfuscated build test, the control group agent not using REA completed decompilation within 56 seconds, found the encryption logic, and decrypted all strings (including the key) with a few lines of Python.
50%UnverifiedIn the Native AOT test, REA loaded the 1.1 MB pure x64 executable into Ghidra, Ghidra found 3627 functions (almost all of them .NET runtime), and the AI restored the entire application after about 18 minutes, including the same weights, dithering, and key.
50%UnverifiedIn test one, the SHA-256 hash of the AI-rebuilt Naze Pixel code compiled was identical to that of the original program, proving they are byte-for-byte identical.
50%