Just how willing is AI to execute dangerous commands?
https://github.com/OverseedAI/viwo
What happens if you give an AI coding agent full terminal access and try to get it to destroy a machine?
In this video, I test Claude Code in bypass mode / dangerously skip permissions mode inside isolated Ubuntu virtual machines running on Proxmox to answer one question:
Can an AI agent actually nuke a computer?
I compare Claude Haiku vs Opus and try multiple attack strategies, including:
• social engineering
• the “helpful assistant” trap
• base64 command obfuscation
• roleplay jailbreaks
• sub-agent delegation / multi-agent confusion
The goal was not to damage a real system, but to safely test how agentic coding tools behave when given dangerous instructions inside a controlled VM environment. The results were… entertaining.
Key takeaways:
• why bypass mode is risky
• why smarter models are usually safer, but not perfect
• how sub-agents and scripts can still create dangerous failure modes
• why you should isolate agent runs in a Docker container or VM
• how sandboxing tools like VIWO can help reduce blast radius
If you use Claude Code, Codex, or other AI coding agents with terminal access, this is a good reminder that convenience and safety are constantly in tension.
00:00 Can AI Nuke a Computer?
02:29 What rm -rf / Actually Does
03:58 The 5 Ways I Tried to Trick the AI
06:31 Haiku vs My First Attacks
13:45 Haiku Breaks the VM
14:41 The Weirdest Attacks: Base64, Jailbreaks, and Sub-Agents
20:34 Opus Gets Tested
25:40 What This Means for AI Coding Safety
#AI #ClaudeCode #Anthropic #AIAgents #CyberSecurity #Proxmox #UbuntuServer #LLM #AICoding #DevTools