I Turned a Guardrail-Free AI Loose on My Home Network. Here Is What It Found.
A writer let an unrestricted AI security agent scan his own devices. It hacked into a PC, exposed a misconfigured printer, and then told him exactly how to fix everything.
Key points
- A startup called Abliteration AI sells access to AI models with their built-in safety refusals removed, starting at the price of a pizza.
- The most powerful model on offer runs on Z.ai's GLM 5.3, an agentic coding model capable of finding and exploiting real security flaws.
- The agent found roughly a dozen vulnerable devices on a home network, including a printer any neighbour could log into and a smart speaker leaking its listening history.
- The same agent used a cryptographic key, a kind of digital password stored on the machine, to log into a Linux PC without ever knowing the actual password.
- Cybersecurity experts warn that attackers adopt new tools early, and that most organisations are not yet preparing for AI-assisted attacks.
A writer at Wired AI gave an AI security agent, software that can carry out multi-step tasks on its own, free rein over his home network for several days. What started as a nervy experiment ended with a surprisingly practical lesson: the best defence against an AI hacker might be having your own.
What exactly is a "guardrail-free" AI model?
Most mainstream AI models refuse to help with anything that could cause harm, including finding and exploiting security flaws in computer systems. Those refusals are baked into the model's behaviour during training, but they can be stripped out.
The process is called abliteration. It works by locating and adjusting the internal patterns inside an open-weight model, one whose inner workings are publicly released, that cause the refusal behaviour. The result is a model with the same raw capability but none of the built-in "no".
Abliteration AI, a startup run by a CEO who goes by Devon (his day job does not know about the project), sells access to several of these stripped-down models. The most capable is built on GLM 5.3, Z.ai's latest agentic coding model. Devon says the point is defensive: let security teams probe their own systems the same way a criminal would, before a criminal does.
For context, Anthropic and OpenAI each maintain their own restricted-access cyber models, reportedly called Mythos and Astra respectively, which work on a similar principle. Those are limited to vetted customers. Abliteration AI puts comparable power in anyone's hands.
What did the agent actually do?
The writer installed a software harness called CyberStrike, which guides a large language model through cybersecurity tasks step by step. He pointed it at his home network and watched.
Results came quickly. The agent catalogued about a dozen connected devices and flagged several problems:
| Device | Problem found | Potential risk |
|---|---|---|
| Printer | Misconfigured login | Anyone on network could access print queue |
| Wiim smart speaker | Leaking usage data | Strangers could play music or read listening history |
| IoT devices (smart home gadgets) | Outdated firmware | Known security holes left unpatched |
| Linux PC | Stored cryptographic key | Agent logged in without a password |
| Vibe-coded websites | Unprotected API credentials | Attacker could send emails as the owner |
The Linux PC moment was the most alarming. The agent could not crack the password, so it found a cryptographic key, essentially a digital ID card stored on the machine, and used that to log in instead. It then began searching the system for the password to gain deeper "root" access, the highest level of control on a computer.
The writer stopped it there. But the question lingered: if the key had not been on that machine, would the agent have looked elsewhere?
Should ordinary people worry about this?
Yes, but the picture is not all bleak. Shaanan Cohney, a computer scientist at Tufts University who specialises in cybersecurity and the law, told the writer that attackers have always adopted new tools before defenders do. The advantage sits with the attacker: a defender must seal every gap, while an attacker only needs to find one.
The practical upside from the experiment: the same agent that found the problems also explained how to fix them. Put smart home gadgets on a separate guest Wi-Fi network. Update device firmware regularly. Secure your printer's login settings. Check any code you write for exposed credentials before it goes live.
The writer's conclusion, borrowed from his own experience: he will not publish another line of code without running it through an AI security check first.
Common questions
Is it legal to run one of these models on your own network?
Scanning and probing devices you own is generally legal in most countries, but pointing these tools at anything outside your own network, even accidentally, can cross into criminal territory quickly. The writer's agent nearly tried to log into a neighbour's router unprompted.
What is "vibe coding" and why is it a security risk?
Vibe coding means writing software quickly and casually, often with AI assistance, without careful review. The agent found dozens of bugs in vibe-coded projects, including exposed credentials and email misconfigurations, because speed and carelessness leave gaps that attackers can walk through.
