AI-Crafted Zero-Day: Google Stops the First Cyberattack Written Entirely by Artificial Intelligence

"The new era is here." That's not a tagline from a sci-fi movie — those are the exact words of John Hultquist, chief analyst at Google's Threat Intelligence Group, after his team confirmed what the cybersecurity world had been dreading: the first zero-day exploit written entirely by artificial intelligence, found in the wild, ready to be deployed against real targets. And the thing is, this isn't some distant theoretical risk anymore. It happened. In May 2026. The AI arms race in cybersecurity just became very, very real.

"The new era is here. We have confirmed the first AI-crafted zero-day exploit in the wild. The barrier between script kiddies and nation-state hackers just collapsed."

If you're running a business, managing IT systems, or simply relying on two-factor authentication to keep your accounts safe — and let's be honest, who isn't? — this story should make you sit up. Not to panic, but to understand. Because understanding this shift is what separates the companies that get hit from the ones that saw it coming.

What Google Actually Found: A Python Script That Shouldn't Exist

In May 2026, Google's Threat Intelligence Group (GTIG) released evidence of something unprecedented: a fully functional zero-day exploit — a Python script — designed to bypass two-factor authentication on a popular open-source web-based system administration tool. But here's what made it different from every exploit that came before it: it was written by an AI.

How did they know? It turns out that AI-generated code leaves fingerprints — subtle but unmistakable patterns that no experienced human hacker would ever produce. The GTIG team identified multiple telltale signs:

Think of it this way: it's like finding a burglary tool that comes with a professionally printed instruction manual, a warranty card, and a customer satisfaction survey. Real criminals don't do that. But AI does — because it was trained on educational content and doesn't know the difference between helping a student learn and helping a criminal attack.

How the 2FA Bypass Actually Worked: The Semantic Logic Error

Here's where it gets really clever — and honestly, a little scary. The vulnerability wasn't a coding mistake in the traditional sense. It wasn't a buffer overflow or a SQL injection. It was a semantic logic error: a place where the developer hardcoded a trust assumption that directly contradicted the authentication enforcement logic.

In plain English? The system's code essentially said: "Always trust requests coming from this internal pathway" while simultaneously saying "Always verify identity before granting access." These two rules contradicted each other, and the AI found the gap. It wrote a script that exploited that contradiction to slip past 2FA completely.

The truth is that this is the kind of vulnerability that human code reviewers have missed for years. It doesn't look like a bug. It looks like a design decision. And that's exactly the type of nuanced logical reasoning that modern LLMs excel at — they can read thousands of lines of code and spot contradictions in the intent of the logic, not just the syntax.

The AI didn't brute-force anything. It didn't need to. It simply read the code, understood the logical contradiction, and wrote a clean exploit. That's more sophisticated than 90% of human hackers.

The North Korean Connection — And What Google's Gemini Wasn't Used For

GTIG made a point of clarifying that Google's own Gemini AI was not used to create this particular exploit. However — and this is a big however — they revealed that North Korean state-sponsored hackers (APT45) had been sending thousands of prompts to Gemini specifically for exploit research. Thousands. Not a handful of curious queries. A systematic, industrial-scale effort to weaponize AI for cyberattacks.

The cybercrime group behind the zero-day had planned mass exploitation — deploying the AI-crafted bypass at scale against organizations worldwide — before Google stopped them. GTIG worked directly with the affected vendor to responsibly disclose the vulnerability and deploy patches before the attack could go live.

And the thing is, this was just the one Google caught. How many others are already out there, written by AI systems with less rigorous monitoring?

The Bigger Picture: When AI Finds 10,000 Vulnerabilities in a Month

If one AI-crafted exploit sounds bad, consider what Anthropic revealed around the same time. Their Claude Mythos Preview AI, through a project called Glasswing, discovered over 10,000 critical vulnerabilities across major open-source projects in just one month.

Ten thousand. In thirty days.

The volume was so overwhelming that open-source maintainers actually asked Anthropic to slow down the disclosures. Of 530 bugs formally reported, only 75 had been patched when the request was made. The maintainers simply couldn't keep up.

Now, here's the thought that should keep every business owner awake: if the good guys found 10,000 vulnerabilities this quickly, the bad guys are doing the exact same thing. Except they're not reporting them — they're exploiting them.

Why This Changes Everything for Your Business

Let's be direct about what this means. The barrier to sophisticated hacking has collapsed. It used to take years of specialized training to find zero-day vulnerabilities and write working exploits. Now? An AI can do it in hours. And it doesn't need to sleep, take breaks, or learn from scratch.

Senator Chuck Schumer warned in May 2026 about AI-enabled hacking threats to hospitals, energy grids, and water infrastructure. But in addition to these headline-grabbing targets, every mid-sized company running web-based admin tools, using standard 2FA implementations, or relying on open-source software in their stack is now potentially exposed.

• If you have a web-based admin panel — it could contain the same type of semantic logic error that this AI exploited
• If you rely on SMS-based 2FA — AI-powered bypass techniques are evolving faster than implementations can keep up
• If you use open-source tools — they likely contain some of those 10,000+ undiscovered vulnerabilities
• If your last security audit was more than 6 months ago — the threat landscape has fundamentally changed since then

PROMPTSPY: The Autonomous AI Malware Already on Phones

As if AI-crafted exploits weren't enough, there's also PROMPTSPY — a newly discovered Android malware that takes AI threats to an entirely new level. This isn't just AI-assisted code. This is malware that uses AI to think, adapt, and make decisions on the victim's device.

PROMPTSPY leverages the Google Gemini API to autonomously navigate victim devices. Unlike traditional malware that follows a script ("go to this folder, copy this file, send it here"), PROMPTSPY uses AI reasoning to:

1. Navigate unfamiliar app interfaces it's never seen before
2. Find and capture biometric data based on contextual understanding
3. Make real-time decisions about what's worth stealing
4. Adapt its behavior when it encounters security measures
5. Autonomously explore device contents without predefined instructions

It's the difference between a burglar who follows a map versus a burglar who can think, improvise, and figure out your house layout on the fly. The second one is infinitely harder to defend against — and that's what PROMPTSPY represents.

Supply Chain Attacks Getting AI-Enhanced Too

In March 2026, a group tracked as TeamPCP/UNC6780 compromised legitimate GitHub repositories — including popular projects like LiteLLM and Trivy — injecting them with a credential stealer called SANDCLOCK. These weren't random repos. They're tools used by developers and security professionals, meaning the attackers specifically targeted the software supply chain to reach high-value organizations.

The implication is clear: even installing "trusted" open-source tools now requires more scrutiny than ever. If attackers can compromise the GitHub repositories of security tools themselves, no part of the software supply chain can be blindly trusted.

"We are witnessing the industrialization of vulnerability discovery. What used to require a team of experts and months of work can now be accomplished by AI in an afternoon. The defenders need to move just as fast."

Protection Strategies: What Actually Works Now

So what do you do? Panic? No. But you absolutely need to upgrade your thinking about security. Here's what works against AI-powered threats:

1. Upgrade to Phishing-Resistant MFA

   SMS-based 2FA and even app-based TOTP codes are increasingly vulnerable. Hardware security keys (FIDO2/WebAuthn) resist the class of logical bypass attacks that AI is finding. If you protect anything of value, hardware keys should be non-negotiable for admin access.

2. Aggressive Patch Management (24-48 Hour Windows)

   When AI can find vulnerabilities in hours and generate exploits in minutes, the old "monthly patch cycle" is a death sentence. Critical patches need to deploy within 24-48 hours. If you can't do that internally, you need a managed service that can.

3. AI-Aware Security Monitoring

   Traditional signature-based detection won't catch AI-generated exploits because each one is unique. You need behavioral monitoring that detects unusual authentication patterns, anomalous API calls, and logic-level bypasses — not just known malware signatures.

4. Audit Hardcoded Trust Assumptions

   The specific vulnerability Google found was a hardcoded trust assumption. Review every place in your systems where authentication can be bypassed through "trusted" internal pathways. AI will find these — it's only a matter of when.

5. Assume Breach, Plan Recovery

   With the volume of vulnerabilities AI is discovering, perfect prevention is impossible. Your backup strategy, incident response plan, and recovery procedures are now more critical than ever. Test them regularly.

The Timeline of AI Threats: How Fast Things Changed

Look at that timeline. That's not a year's worth of escalation. That's three months. The acceleration is staggering, and it means that any security posture built on last year's assumptions is already outdated.

What Makes You Smarter Than the Next Target

Here's the good news, and the reason you should feel empowered rather than paralyzed: the organizations that get hit are overwhelmingly the ones that didn't adapt. They kept SMS 2FA when hardware keys existed. They patched monthly when weekly was necessary. They relied on antivirus signatures when behavioral monitoring was available.

You're reading this article. You now understand that AI-generated exploits are real, that they target logical contradictions rather than simple coding mistakes, and that the volume of vulnerability discovery has gone exponential. That understanding — that awareness — is genuinely your first line of defense.

The next step is action. And the truth is that for most growing businesses, the complexity of AI-aware security requires professional support. Not because you're incapable, but because the speed of this arms race demands full-time attention from people who do nothing else.