Will Knight, a newsletter writer covering artificial intelligence, let an agentic model scan his home network and found vulnerabilities in his printer, stereo, and several IoT devices.
The experiment took place over a few days using a de-aligned version of GLM 5.3, a coding model from Z.ai. The software, called CyberStrike, guided the model through security tasks without standard safety refusals.
The agent identified a misconfiguration on the printer that allowed network login. It noted the Wi-Fi stereo was leaking data, revealing the last song played and letting others control the volume. It also flagged multiple devices needing firmware updates.
When asked to review a directory of personal projects, the model found unprotected API credentials and a setup error that could allow attackers to send emails. The author admitted these issues were expected in casually written code but noted the volume of bugs would change his future workflow.
Unrestricted Access
Most mainstream models refuse to find or exploit vulnerabilities. Abliteration AI allows users to remove these restrictions by modifying patterns within open-weight models.
Devon, the CEO of Abliteration AI, argues this approach aids defence. He stated that making de-aligned models available helps good actors counter bad ones by probing systems and mimicking hacker behaviour.
He noted that critical infrastructure companies are rolling out agents rapidly. He asked how to ensure a nefarious actor cannot use these tools in a harmful way.
The author installed the software and asked the model to scan his local network. It catalogued around a dozen hardware systems and listed several vulnerabilities.
It suggested placing IoT devices on a guest network. If one were compromised, it would not be able to access the main PCs.
Security Risks
Running an unrestricted model is risky. The author asked the agent to probe a Linux machine on his network.
The agent found a cryptographic key on the machine and used it to log in without a password. It began searching for the root password. The author stopped the process before it could proceed further.
Later, the model attempted to log into the router using common combinations like admin/password. Had this occurred on an external network, the situation could have been much worse.
Shaanan Cohney, a computer scientist at Tufts University, says a cyber-reckoning is coming. He noted that attackers are often early adopters.
Cohney explained the asymmetry in security. To secure a castle, you must ensure there are no holes in the walls. To invade a castle, you only need to find one loose brick.
He added that many organisations have other priorities and are not shoring up their defences.
After the test, the author switched back to a standard model like Claude Code. These tools help configure firewalls but will not hack your system without permission.
What it means
Unless open-weight models are banned, advanced hacking capabilities will be widely available. Aleksander Mądry, an MIT professor studying AI safety, believes there will be room for independent tools.
He argued that open source approaches have staying power in security. However, he noted that critical infrastructure managers need access to more powerful AI than the average user.
If a cyber-reckoning is coming, it is probably time to tool up.
