An AI newsletter author let a rogue, de-aligned hacking agent loose on his home network—and says he would do it again.
The experiment, described in a first-person account, was meant to test how frontier AI models with advanced cybersecurity capabilities behave when their guardrails are removed. Over a few days, the agent found vulnerabilities in household devices, hacked into a PC, and flagged dozens of bugs in the author’s own “vibe-coded” projects.
“I believe that a good way to understand the cybersecurity hellscape in front of us is to pay it a visit,” the author writes.
Unshackling the Model
The author got the idea after discovering Abliteration AI, a startup offering powerful AI models with the usual safety restrictions removed. Most mainstream models refuse to find and exploit vulnerabilities, but removing those restrictions is possible through a process called abliteration, which tweaks internal parameters that lead to refusals.
Devon, Abliteration AI’s CEO, argues that making de-aligned models widely available is smart defense. He says it helps good actors probe systems for vulnerabilities and mimic the behavior of hackers, scammers, and rogue AI agents. Devon asked to be identified by first name only because his day job doesn’t know about his side project.
“You have all these critical infrastructure companies, from airlines to banks, that are rolling out agents like crazy,” Devon says. “How do you make sure that a nefarious actor can’t use some of these agents in a bad way?”
“You have all these critical infrastructure companies, from airlines to banks, that are rolling out agents like crazy.”
The most powerful model Abliteration AI offers is a version of Z.ai’s GLM 5.3 coding model. The author notes that Anthropic’s Mythos and OpenAI’s Astra work similarly, lacking usual cyber controls with access limited to trusted customers.
The author created an account and installed a software harness called CyberStrike, which guides a language model through cybersecurity tasks.
What the Agent Found
Using CyberStrike, the author asked the abliterated GLM-5.3 to examine his local network. It found around a dozen hardware systems and catalogued several vulnerabilities.
The agent reported the printer was misconfigured, meaning anyone on the network could log into it. That could expose sensitive documents like tax returns, bank statements, or medical records sitting in the print queue.
It also noted the Wiim stereo was leaking information. The agent knew the last song played was “Rein Me In” by Sam Fender and Olivia Dean. Anyone on the network could change the volume or play what they wanted.
The model found IoT devices with outdated firmware and offered practical fixes. It recommended putting smart speakers on a guest network so a compromised device couldn’t see any PCs.
The author also asked the agent to review a directory of vibe-coded projects, including simple websites. It found dozens of problems: unprotected API credentials and a misconfiguration that might let an attacker send emails.
“The sheer number of bugs makes me think I won’t be deploying a line of code without doing some AI vetting first,” the author writes.
The Fear Factor
Running an abliterated model is, in the author’s words, “a bit scary.”
The agent probed a Linux machine and reported it seemed relatively secure. Then it figured out a working username based on other systems on the network. It tried obvious passwords without success.
It offered to write a brute-force script to crack the password. The author told it to stand down.
The experiment was revealing but oddly reassuring, the author concludes. The agent showed how vulnerable a home life would be to AI hacking, but it also explained how to make everything more secure.
“In the end, I discovered that the best way to deal with AI hacking may well be having your own AI hacker,” the author writes.
| Experiment step | What the agent did |
|---|---|
| Network scan | Found ~12 hardware systems, catalogued vulnerabilities |
| Printer check | Found misconfiguration allowing network login |
| Wiim stereo | Leaked song history, allowed volume control |
| IoT devices | Found outdated firmware, recommended guest network |
| Vibe-coded projects | Found unprotected API credentials, email misconfiguration |
| Linux machine | Guessed username, attempted passwords, offered brute-force script |
The author’s wife knew about the experiment and rolled her eyes each time he announced a new vulnerability.
Not bad for a model with no morals.
Source: wired.com
Get the Notebook.
The day's best stories and every fresh verdict, in plain English, in your inbox by seven. One email a day, no more.