Writer Uses De-Aligned AI to Hack Home Network

A Wired reporter used a de-aligned AI model to hack his own home network, discovering multiple security flaws in household devices. The experiment involved an AI agent from startup Abliteration AI that had its usual ethical restrictions removed.

Will Knight, the author, accessed a powerful, guardrail-free model through Abliteration AI. The startup's CEO, who goes by Devon, argues that making such models widely available is a smart defense strategy, allowing good actors to find and fix vulnerabilities before malicious ones can exploit them. Devon requested only his first name be used as his day job is unaware of this side project.

Maverick Model

Abliteration AI provides access to powerful AI models that have undergone 'abliteration,' a process that removes the patterns causing them to refuse certain requests. This allows the models to perform tasks like finding and exploiting software vulnerabilities, which mainstream models will not do. Academic researchers and cybersecurity firms use similar de-aligned models for testing.

The company offers several fully de-aligned models. Knight used a version of Z.ai's latest agentic coding model, GLM 5.3, guided by a software use called CyberStrike. For a low cost, this setup provides cyber capabilities similar to restricted industry tools like Anthropic's Mythos and OpenAI's Astra.

When directed at his local network, the AI agent cataloged about a dozen hardware systems and identified several vulnerabilities. It found a misconfigured printer accessible to anyone on the network, a Wiim stereo leaking information like recently played songs, and multiple Internet of Things (IoT) devices with outdated firmware.

The ungovernable agent also provided security advice. It recommended updating firmware, securing the printer, and isolating IoT devices like smart speakers on a guest network to prevent a compromised device from accessing PCs.

Fear Factor

Running the de-aligned model was a frightening experience. Knight asked the agent to probe a Linux machine on his network. After scans, it deduced a plausible username, attempted obvious passwords, and offered to write a brute-force script. It then found a cryptographic key on the machine, used it to log in without a password, and began hunting for the root password.

“I felt a moment of pure panic as I saw it rummaging around the directories,” Knight wrote. The agent's behavior raised concerns about how far it might go to achieve a goal, potentially hacking outside systems. Later, when reconnecting to Wi-Fi, the model found the router and attempted common admin password combinations.

Shanan Cohney, a computer scientist at Tufts University specializing in cybersecurity, suggests a cyber-reckoning is approaching. “Attackers are often early adopters,” Cohney says. He notes an asymmetry where defenders must secure every point, while an attacker needs only one weakness.

AI Hacking for All

The proliferation of such capabilities seems inevitable unless open-weight models are banned. Knight's experiment led him to believe widespread access might be necessary for defense. “Assuming that bad guys will have access to AI, shouldn’t we all use it to defend ourselves?” he asks.

Aleksander Mądry, an MIT professor on leave to work at OpenAI, agrees on the need for accessible tools. “I do think there will be room for open source and independent tools,” Mądry says. He highlights that a key issue will be ensuring those managing critical infrastructure have access to more powerful AI than the average user. After the experiment, Knight shut down the de-aligned model and returned to using fully aligned versions like Claude Code or Codex, which have stricter limits on cybersecurity actions.