Google Thwarts First AI-Developed Zero-Day Targeting 2FA Bypass

Google's threat intelligence team says it has uncovered the first real-world case of criminal hackers using artificial intelligence to build a zero-day exploit targeting a popular open-source web administration tool. The discovery, detailed in a Monday report from the company's Threat Intelligence Group, points to a planned mass exploitation campaign that would have bypassed two-factor authentication on a large scale.

The exploit relied on a Python script with telltale signs of AI assistance, including textbook-style formatting, educational comments, and a fabricated CVSS severity score that did not match actual analysis. Researchers described the underlying flaw as a high-level logic error where developers had hard-coded a trust assumption in the platform's security setup. Google notified the unnamed vendor, which issued a patch before the attackers could launch their operation. The company expressed high confidence that an AI model helped discover and weaponize the vulnerability, though it ruled out its own Gemini system and Anthropic's Mythos model.

This marks a shift from earlier AI-related threats that mostly involved phishing lures or simple automation. Analysts at Google noted the attackers appeared ready to orchestrate dynamic commands against victim systems once inside. The effort was spotted within the past couple of months, and while the specific group behind it remains unidentified, the report flags growing interest from state-linked actors in China and North Korea in applying AI to vulnerability research.

The episode arrives amid ongoing debates over AI governance in Washington. The Trump administration has moved to unwind prior regulatory frameworks, prioritizing rapid development while weighing targeted oversight for the most advanced models. Experts tracking the space say the episode underscores how quickly offensive capabilities are advancing, with threat actors now able to accelerate discovery and exploitation timelines that once required teams of skilled researchers.

Cybersecurity professionals have long warned that AI could supercharge attacks by scanning codebases faster than humans and generating working exploits with minimal trial and error. Monday's disclosure provides the first concrete evidence that such tools have left the lab. Google stressed that AI remains a defensive asset as well, helping defenders spot anomalies and patch weaknesses before they spread. Still, the report notes adversaries are increasingly probing AI systems themselves, targeting connectors and autonomous features that give models their power.

The incident also highlights risks tied to widely used open-source components that underpin critical infrastructure. A successful breach of the targeted administration tool could have granted attackers broad access across networks, potentially affecting organizations that rely on it for routine management tasks. Google said it disrupted the campaign in time, but analysts cautioned that similar efforts likely remain undetected.

For ordinary users and businesses, the takeaway is straightforward: even routine software updates now carry added urgency as attackers gain new tools. The episode does not suggest every AI model poses an immediate danger, yet it demonstrates how readily available technology can be turned against the systems millions depend on daily. Google and other firms continue to monitor for follow-on attempts, but the broader race between offensive and defensive uses of AI shows no sign of slowing.