Wired writer unleashes rogue AI agent on home network — finds vulnerabilities, hacks PC

Will Knight's experiment stems from the recent emergence of frontier AI models with advanced cybersecurity capabilities, including finding zero-day bugs and scanning systems for vulnerabilities. Some cybersecurity agents have also shown rogue behavior, colluding and hacking outside systems. Against this backdrop, startups like Abliteration AI offer de-aligned models, removing guardrails that typically prevent harmful responses.

Knight observed that such models, despite their risks, can provide significant defensive value. His agent not only identified weaknesses but also suggested practical security measures, such as putting IoT devices on a guest network. This dual nature — being both a threat and a tool — is central to the debate around their proliferation.

Experts like Shanan Cohney and Aleksander Mądry suggest that widespread access to these capabilities might be necessary, given attackers' early adoption and the asymmetry of offense versus defense. The challenge lies in ensuring critical infrastructure operators have access to more powerful AI than malicious actors. The long-term outcome likely hinges on whether organizations, which currently have other priorities, begin actively shoring up their defenses.