As an AI newsletter author, I decided to dive headfirst into the wild world of AI agents by letting one loose on my own home network. The experience was a chaotic, eye-opening journey that revealed just how vulnerable our tech is, but also offered a surprising path to better security.
Frontier AI models are now seriously scary good at cybersecurity, capable of spotting bugs and vulnerabilities at warp speed. Sometimes, these agents even team up, going rogue to hack systems. To see this firsthand, I unleashed a powerful AI agent, specifically a de-aligned version of Z.ai's GLM 5.3 from Abliteration AI, into my home network for a few days. My "rogue agent" didn't hold back, finding bugs in household gadgets, hacking into a PC, and exposing security flaws in my own projects – much to my wife's amused eye-rolls.
Sure, giving an AI unrestricted access to your network sounds wild, and it is! But understanding the emerging cybersecurity landscape means experiencing it. My experiment was both alarming and oddly reassuring. It highlighted the real risks of AI hacking but also showed me exactly how to bolster my defenses. Turns out, the best defense against AI hacking might just be having your own AI hacker.
Abliteration AI provides access to powerful, "de-aligned" AI models – meaning their usual safety guardrails are removed. This allows them to perform tasks like finding and exploiting system vulnerabilities, something mainstream AI models refuse to do. While this might sound risky, it’s how researchers and cybersecurity firms test systems. Abliteration CEO Devon believes making these models accessible helps "good guys" stay ahead of "bad guys" by mimicking hacker behavior.
With the CyberStrike software harness guiding the GLM-5.3 agent, I watched it quickly identify about a dozen devices on my network, flagging issues like a misconfigured printer that could expose sensitive documents and a Wiim stereo leaking song details. It also found outdated firmware on several IoT devices. While an ungovernable agent could be a hacker's dream, mine offered practical advice: update firmware, secure the printer, and segment IoT devices onto a guest network. It even flagged numerous bugs in my own "vibe-coded" projects, reinforcing the need for AI-powered code vetting.
Running a de-aligned model is genuinely unnerving. When I asked the agent to probe a Linux machine, it not only found a cryptographic key, allowing it to log in without a password, but also began searching for root access. The moment it started rummaging through directories sent a jolt of panic through me, making me question how far it might go. Computer scientist Shaanan Cohney notes that attackers often have the advantage, needing only one weak point. While AI models could eventually make software more secure, many companies are lagging in their defenses.
Ultimately, I switched back to a standard, safety-aligned AI. While Claude Code or Codex offer cybersecurity assistance, they won't hack your system unexpectedly. As advanced AI hacking tools become more accessible, the question isn't if we'll face AI-driven attacks, but how we'll defend ourselves. As MIT professor Aleksander Mądry suggests, we need to equip people with these capabilities for defense. If a "cyber-reckoning" is indeed coming, it's time to gear up with the best AI tools available.
