Researcher Removes AI Guardrails. It Hacks Their Own PC. The Method Matters, Obviously.
The author created an Abliteration AI account and installed a software harness called CyberStrike to guide a large language model through cybersecurity tasks. With safety guardrails removed, the agent found vulnerabilities in household devices, hacked into a PC, and identified dozens of problems in vibe-coded projects, including unprotected API credentials and a misconfiguration that could let attackers send emails. The AI also explained how to secure everything it broke.
This demonstrates the concept of dual-use AI systems. The same capability that finds vulnerabilities can patch them. The underlying principle is that offensive security testing and defensive hardening are two applications of the same analytical mechanism. Anyone building casual software should understand that automated agents can now audit code faster than most humans can write it.
The author used Abliteration AI and the CyberStrike software harness to run an open-source model through cybersecurity tasks on their own home network and personal projects.
- Sign up for a free account at Abliteration AI and explore their interface to see how guardrail-removed models function.
- Take any small coding project you have, even a simple website, and paste its code into a capable AI assistant like Claude or ChatGPT with the prompt "Audit this code for security vulnerabilities and list every issue you find."
- Review the AI's output and implement at least one suggested fix. You will likely discover problems you never considered, which is the entire point.