Google's Threat Intelligence Group has confirmed what security researchers have been dreading: a real, in-the-wild zero-day exploit that was almost certainly written by an AI system, not a human. The target was a popular open-source web administration tool, the exploit was designed to bypass two-factor authentication, and a prominent cybercrime group was gearing up to use it for mass exploitation before Google's researchers spotted and stopped it.

The tell? Artefacts baked into the Python exploit script that were inconsistent with how human developers write code but consistent with how large language models generate it. It's the first time Google — or anyone credible — has publicly confirmed AI-assisted exploit development detected in active attack infrastructure, rather than in a research lab proof-of-concept.

Why this is a milestone, not just another vuln story.

We've known for a while that attackers were experimenting with AI to speed up their workflows — drafting phishing lures, summarising vulnerability disclosures, accelerating reconnaissance. Using an LLM to actually generate a functional, novel zero-day exploit and deploy it operationally is a different category of threat. It compresses the time between "vulnerability exists" and "working exploit is in someone's hands."

Think of it this way: historically, turning a newly-discovered software flaw into a reliable, weaponised exploit takes skill, time, and iteration. That gap — sometimes weeks, sometimes months — is where defenders patch, detect, and respond. AI-assisted exploit development starts closing that window in a way that's qualitatively different from, say, buying an exploit on a darknet forum.

The 2FA angle is particularly pointed. The targeted web administration tool presumably has 2FA deployed as a security control. Bypassing that in a mass exploitation campaign would have given attackers authenticated access to potentially thousands of servers with no credential theft required. For organisations running this tool, MFA alone wouldn't have saved them.

What we don't yet know is which administration tool was targeted, which cybercrime group was behind it, and whether the exploit has circulated more widely since Google's disclosure. The lack of specifics is frustrating but not unusual — responsible disclosure timelines often mean details stay vague until patches are shipped.

The defender's takeaway is uncomfortable. If AI can generate viable zero-days, then the asymmetry between attackers and defenders gets worse. Attackers already have speed and surprise on their side. AI giving them an exploit-generation capability at scale means security teams need to lean harder into detection and response — because prevention alone, at this pace, isn't going to hold.

This is also a reminder that the AI-as-attack-tool story isn't theoretical anymore. It isn't red-team researchers demonstrating what could happen. It happened. Google caught it this time. The question is how many times it won't be caught.

For Australian organisations running open-source web admin infrastructure — cPanel, Webmin, and similar tools are widely deployed in Australian hosting environments — this is a prompt to audit your exposure, ensure you're on current versions, and check whether your monitoring would actually catch authenticated-but-anomalous post-login behaviour if a bypass like this succeeded.