WIRED test shows how an unguarded AI agent exposed home-network weaknesses

A WIRED writer testing an open-source artificial intelligence model with its safety guardrails removed found vulnerabilities across a home network, accessed a Linux PC using a cryptographic key and identified weaknesses in several software projects.

The test used CyberStrike with an altered version of the GLM-5.3 model obtained through Abliteration AI. The model reportedly identified about a dozen hardware systems, including a printer whose settings allowed network access, a WiiM stereo that exposed information and internet-of-things devices running outdated firmware.

The agent also found unprotected API credentials and an email-sending misconfiguration in projects the writer described as casually coded. On the Linux machine, it identified a working username, accessed the system using a cryptographic key and began searching for a password that could provide root access. The source does not establish that the machine was fully compromised.

In a later test, the agent detected the home router and attempted common login combinations. The writer stopped some actions, and the experiment did not establish whether the model could have accessed systems outside the local network.

WIRED reported that the agent also provided defensive recommendations, including updating firmware, securing printers and placing IoT devices on a guest network. The findings come from a first-person experiment and have not been independently verified in the source material.

The test illustrates the tension around broadly available cyber-capable AI. Tufts University computer scientist Shaanan Cohney said attackers often adopt new tools early, while MIT AI-safety researcher Aleksander Mądry argued that open-source and independent systems could have a lasting role in security. Both the potential defensive benefits and the risks depend on how organisations prioritise protection.