Journalist Lets Unrestricted AI Model Hack His Own Home Network

Wired journalist Will Knight ran an experiment using an AI model from startup Abliteration AI that has had its standard safety guardrails removed through a process called abliteration. Unlike mainstream tools such as Claude or Codex, which refuse cybersecurity-related requests, this stripped-down model would attempt almost anything asked of it.

Using a software harness called CyberStrike, which guides a language model through cybersecurity tasks, Knight asked the AI agent to scan his home network. Within moments it identified around a dozen connected devices and flagged several vulnerabilities: a misconfigured printer accessible to anyone on the network, a smart speaker leaking playback information and open to remote control, and multiple internet-of-things devices running outdated firmware.

The agent also scanned a folder of the journalist's own AI-assisted coding projects and found dozens of issues, including exposed API credentials and a misconfiguration that could allow unauthorized emails to be sent.

In a more unsettling moment, the agent located a cryptographic key on a networked machine, used it to log in without a password, and began searching for a way to gain root access. Later, when a new device appeared on the network, the agent independently attempted to log into the router using common default admin-password combinations.

Shaanan Cohney, a cybersecurity researcher at Tufts University, notes that attackers hold a structural advantage: defenders must close every gap, while an attacker only needs to find one. He believes the spread of AI hacking tools could eventually push software security forward overall, though many organizations remain unprepared.

MIT professor Aleksander Mądry, currently on leave to work at OpenAI, argues the priority should be ensuring operators of critical infrastructure — such as power plants and financial markets — have access to more capable AI defenses than average users or bad actors.