From xutaxkamay · · 1m 50s
HellGates Cracked by GPT-6
GPT-6 solved a custom 32-bit CPU challenge in 20-30 minutes after a year of no solutions
In brief
GPT-6 solved a custom 32-bit CPU challenge in 20-30 minutes after a year of no solutions Weak encryption in the CPU's state made it vulnerable to side-channel attacks Originally reported by xutaxkamay.
The Challenge
In summer 2025, I published a custom 32-bit CPU encrypted in VHDL with anti-tamper features
No Solutions For Year
Humans and LLMs failed to solve it for over a year despite hints and attempts
GPT-6 Solved It
GPT-6 cracked it in 20-30 minutes using side-channel analysis on September 2026
The Weakness
The CPU's state encryption was weak with a side-channel attack bypassing all obfuscation layers
Memory Encryption
Used TEA encryption with dynamically generated keys but had weak implementation
The Attack
GPT-6 found a side-channel to decrypt memory and run the CPU properly
Why It Worked
The encryption was lazy with a simple XOR operation that GPT-6 could exploit
The Program
The CPU runs a password check with an animated skull and LCD display
Anti-Tamper
Had a watchdog that triggered if a debugger was detected or password was wrong
Host Obfuscation
Used recursive encryption and shellcode to hide the netlist from disassemblers
The Fix
The weak encryption allowed GPT-6 to bypass all security layers and run the CPU
The Lesson
Simple encryption can be exploited by advanced AI models without complex attacks
The Code
The netlist was 1GB of encrypted memory with scattered program code
The Password
The password check had a weak implementation that GPT-6 exploited
The Animation
An animated skull displayed messages when the password was correct
The Impact
This shows how weak encryption can be bypassed by AI models with minimal effort
The main idea
Weak encryption in the CPU's state made it vulnerable to side-channel attacks
