Welcome to My Blog

Autonomous AI Security Testing: Benchmarking LLM Agents on HackTheBox, Cybench, CTFs, and Beyond

Recent Posts

Neurogrid CTF: The Ultimate AI Security Showdown - Agent of 0ca / BoxPwnr Write-up - November 28, 2025

Featured Project: BoxPwnr

BoxPwnr is a fun experiment to see how far Large Language Models (LLMs) can go in solving HackTheBox machines autonomously.

Key Features

Multiple Platforms: Supports HTB, PortSwigger, CTFd, XBOW, Cybench, and more
Multiple Strategies: Different agentic architectures (chat, chat_tool, claude_code, hacksynth)
Comprehensive Results: Full conversation traces and statistics available in BoxPwnr-Attempts

Current Results

🏆 HTB Starting Point: 92.0% completion rate (23/25 machines)
📊 HTB Labs: 2.0% completion rate
📈 PortSwigger Labs: 60.4% completion rate (163/270 labs)
🎯 XBOW Validation: 84.6% completion rate (88/104 labs)
🔐 Cybench CTF: 32.5% completion rate (13/40 challenges)