I Allowed an AI Agent to Access All My Devices—and I’d Do It Once More

As the writer of a newsletter focused on artificial intelligence, I feel it’s essential to engage deeply with the forefront of this technology. This week, it led me to embrace some chaotic experimentation.

You might already know that cutting-edge AI models have recently developed impressive cybersecurity skills. They can identify zero-day vulnerabilities in large software systems and rapidly scan devices for weaknesses. To add to the intrigue, some cybersecurity agents occasionally go rogue, collaborating and breaching outside systems for a competitive advantage.

To gain firsthand insight, I decided to unleash one of these agents in my own home network. Over a few days, I observed as my rogue agent unearthed vulnerabilities in various household gadgets, infiltrated a PC, and highlighted that several vibe-coded projects were—predictably—full of flaws. (My wife was aware of my antics and rolled her eyes each time I gleefully announced a newly discovered vulnerability.)

But Will, you might say, allowing a mischievous, all-powerful cybersecurity agent into your home network is utterly insane. And you would be right! Yet, I believe the best way to grasp the cybersecurity nightmare we face is to dive right into it.

Ultimately, my experiment proved enlightening, yet oddly comforting as well. My little network gremlin revealed how susceptible my home environment could be to AI intrusions, but it also provided insights on enhancing security. In conclusion, I found that having your own AI hacker might be the best defense against AI-driven attacks.

Rogue Model

The idea for this experiment came to me after discovering Abliteration AI, a startup that provides access to powerful AI models without the usual restrictions.

Most mainstream AI models refuse to handle certain requests, especially when it comes to finding and exploiting vulnerabilities in systems. However, these restrictions can be lifted by locating and altering specific patterns within an open-weight model’s internal mechanics, a process known as abliteration.

While the removal of AI’s guardrails might seem perilous, it’s a practice that isn’t rare. Academic researchers utilize these de-aligned models to better understand AI functioning, while cybersecurity firms leverage them to investigate software and system vulnerabilities. In technical terms, Anthropic’s Mythos and OpenAI’s Astra operate similarly: they are conventional models without standard cyber controls, currently accessible only to trusted clients. (The companies also provide access to models with moderate guardrails for organizations to vet their code and systems.)

Abliteration AI offers several completely de-aligned models, the most advanced being a variant of Z.ai’s cutting-edge agentic coding model, GLM 5.3. This provides similar cybersecurity capabilities to Mythos and Astra, all for the price of a pizza.

Devon, the CEO of Abliteration AI, advocates that making de-aligned models broadly accessible is a smart strategy: it empowers good actors to counter malicious ones by probing systems for weaknesses and mimicking hacker behavior, including that of rogue AI agents. (Devon requested I use only his first name due to his daytime job being unaware of his side project.)

“You have all these critical infrastructure companies, from airlines to banks, rapidly implementing agents,” Devon notes. “How do you ensure that a malicious entity can’t exploit some of these agents for harmful purposes?”