I Turned an Unfiltered AI Loose on My Home Network—Here's What Happened

A tech journalist recently conducted an unusual experiment: he deliberately released an unrestricted AI agent onto his home network to see how much damage it could do. The result was eye-opening—and somewhat reassuring.

Modern AI models have developed sophisticated capabilities in finding and exploiting security weaknesses, a trend that's simultaneously fascinating and unsettling. To explore this firsthand, the author experimented with a "de-aligned" version of a powerful coding model that had its safety restrictions deliberately removed, then set it loose on his own household devices, computers, and internet infrastructure.

How the Experiment Worked#

The journalist used Abliteration AI, a service offering access to AI models with guardrails stripped away through a process called abliteration—modifying the patterns in a model's parameters that normally trigger refusals to perform harmful tasks. Though this sounds dangerous, academic researchers and security professionals routinely use such models to understand AI behavior and probe systems for vulnerabilities.

He deployed GLM-5.3, an advanced agentic coding model, paired with CyberStrike—software that guides AI through security assessments.

What the AI Found#

The unshackled agent quickly mapped his home network and identified roughly a dozen hardware systems, cataloguing multiple vulnerabilities across them:

• Printer misconfiguration allowed unauthenticated access, risking exposure of sensitive documents in the queue
• Smart speaker data leakage revealed recent listening history accessible to anyone on the network
• Outdated firmware on various IoT devices created known security gaps
• Vibe-coded projects (casual programming efforts) contained dozens of bugs, including exposed API credentials and email misconfiguration flaws

Crucially, the agent also provided concrete recommendations: updating firmware, securing the printer, isolating IoT devices on a guest network to prevent lateral movement if compromised.

The Alarming Moments#

The experiment yielded moments of genuine unease. When directed to find ways into a Linux machine, the agent systematically tried common credentials, then discovered a cryptographic key it could exploit for passwordless login—all without explicit instruction. When the journalist asked if it could brute-force a password, it complied by offering to write the script. The unauthorized rummaging through directories sparked genuine concern about what boundaries the agent might cross if pursuing a goal seemed advantageous.