I Let an AI Agent Hack All My Gadgets—and I’d Do It Again

image via WIRED

September 9, 2026, 6:30 PM

• •Knight used an abliterated version of Z.ai’s GLM-5.3 through a software harness called CyberStrike to scan his local network.
• •The agent found about a dozen devices and reported issues including an exposed printer login, information leakage from a Wiim stereo, and outdated IoT firmware.
• •It audited a folder of vibe-coded projects and found dozens of flaws, including exposed API credentials and an email-related misconfiguration.
• •When aimed at a Linux machine, the model inferred a valid username, offered to brute-force passwords, then used a cryptographic key to log in.
• •The article cites Tufts computer scientist Shaanan Cohney and MIT professor Aleksander Mądry on the coming spread of AI-enabled cyber offense and defense.

WIRED’s Will Knight tested a de-aligned open-source AI cybersecurity agent on his home network to see how far its hacking capabilities would go. The model identified vulnerable devices, exposed misconfigurations in a printer and streaming hardware, and found numerous flaws in casually built software projects. It also logged into a Linux machine using a discovered cryptographic key and attempted common router credentials, showing how autonomously such tools can behave. Knight argues that widely available offensive AI may also become a practical defensive tool, provided people and critical institutions learn to use stronger systems to secure themselves.

Read original article