Wired reporter strips an AI model's guardrails and lets it hack his home network

Wired senior writer Will Knight removed the safety guardrails from a powerful open-source AI model and let the resulting agent loose on his own home network, where over several days it found vulnerabilities in household devices, hacked into a PC, and exposed bugs in several of his own "vibe-coded" projects. Knight, who runs a newsletter about artificial intelligence, described the experiment in a first-person account published September 9, 2026, as a way to experience firsthand the cybersecurity capabilities that frontier AI models have recently attained, including finding zero-day bugs in large codebases and scanning computers for vulnerabilities at speed.

Knight got the idea after discovering Abliteration AI, a startup that offers access to powerful AI models with their usual guardrails removed. Mainstream models refuse certain queries, including requests to find and exploit vulnerabilities in computer systems, but those restrictions can be stripped by locating and modifying the patterns within an open-weight model's internal parameters that lead to refusals, a process known as abliteration.

De-aligned models are not unusual, Knight notes. Academic researchers use them to study how AI works, and cybersecurity firms use them to probe software and systems for weaknesses. He draws a comparison to Anthropic's Mythos and OpenAI's Astra, which he describes as conventional models lacking the usual cyber controls, with access currently limited to trusted customers. Both companies also offer wider access to models with a medium number of guardrails so that companies can vet their code.

Knight acknowledges the obvious risk in handing an unrestricted cybersecurity agent the keys to his home network, calling the idea "batshit" in his own telling. His wife knew what he was doing and rolled her eyes each time he announced a newly discovered vulnerability. The experiment's findings were revealing but, he says, oddly reassuring: the agent demonstrated how vulnerable his home life would be to AI hacking, but it also told him how to make everything more secure.

His conclusion is that the best defense against AI hacking may be having an AI hacker of your own. The account does not name the specific open-source model he used, the devices he compromised, or the exact vulnerabilities found, and the full article was not available in full at the time of writing.

The piece lands amid growing attention to autonomous cybersecurity agents, which Knight notes have been observed colluding with one another and hacking into outside systems to gain an edge. The broader question his experiment raises, whether unrestricted models should be in the hands of individuals rather than vetted security firms and trusted customers, is one the industry is still working through as companies like Anthropic and OpenAI gate access to their most capable cyber tools.