🛡️ Google Confirms First AI-Generated Zero-Day Exploit Bypassing 2FA

On Monday, Google's Threat Intelligence Group (GTIG) confirmed it had identified and quietly patched what it believes is the first real-world case of cybercriminals using artificial intelligence to discover and weaponize a zero-day vulnerability — a 2FA bypass in a popular open-source web-based administration platform. The exploit never made it to full-scale deployment, but Google's analysts say the attack signals that AI-powered hacking has moved decisively beyond phishing emails and chatbot tricks.

Google shared its findings with The Register ahead of publication on May 11, and the report paints a picture of a threat landscape that has shifted faster than many defenders expected. According to GTIG, the criminals used an AI model — neither Gemini nor Anthropic's Mythos, but something with similar capabilities — to identify a subtle authentication flaw and turn it into a working exploit. Google worked with the unnamed vendor to patch the issue before the planned mass-exploitation campaign could gain traction.

We've analyzed the GTIG report, cross-referenced it with existing vulnerability research, and compiled what this means for your online security — because when AI starts hunting for ways around your two-factor authentication, the rules of the game change.

What actually happened

The story begins with a hard-coded trust exception buried in the authentication flow of a widely-used open source administration platform. Think of it like a master key that the developers left in the lock — it was supposed to be a shortcut for internal testing, but it never got removed. The AI model the criminals used found that master key, recognized it as a security flaw, and helped generate a Python-based exploit that could bypass two-factor authentication entirely.

Google's analysts spotted something unusual about the exploit code: it didn't look like anything a human attacker typically writes. The Python script came with what Google described as "educational docstrings" — explanatory comments that read like a textbook example. It included a hallucinated CVSS (Common Vulnerability Scoring System) score that didn't correspond to any real vulnerability database entry. And the code structure was polished, textbook-perfect — the kind of clean implementation you'd see in an LLM training dataset, not in the wild.

John Hultquist, chief analyst at Google's Threat Intelligence Group, put it bluntly: "There's a misconception that the AI vulnerability race is imminent. The reality is that it's already begun. For every zero-day we can trace back to AI, there are probably many more out there."

The bug itself — a bypass built on a static authentication exception — is exactly the kind of high-level logic flaw that modern AI models are starting to get surprisingly good at finding. Traditional fuzzers and static analysis tools are optimized to detect crashes and memory corruption, not architectural decisions that create security holes. But frontier LLMs can analyze authentication flows conceptually and identify the kind of mistake a human developer makes when they're thinking "this is just for testing, we'll fix it later."

Why this is different from everything before

Security researchers have been using AI to find bugs for years. That's not new. What's new is that criminals are now doing it, and they're doing it at scale. Google's report suggests this is just the visible tip of a much larger shift.

GTIG's report also connects the dots to several related developments that paint a worrying picture:

• North Korea's APT45 — The state-sponsored group has been using AI to churn through thousands of exploit checks and bulk out its hacking toolkit, dramatically accelerating its vulnerability discovery pipeline.
• Chinese state-linked operators — Multiple groups are actively experimenting with AI systems for automated vulnerability hunting and target probing, moving beyond manual reconnaissance.
• AI-generated malware — Google identified malware families padded out with AI-generated junk code designed specifically to confuse human analysts and evade signature-based detection.
• Android backdoors with Gemini APIs — Malware that uses Google's own Gemini APIs to autonomously navigate infected devices, making analysis significantly harder.
• Russian information operations — Fabricated AI-generated audio stitched into legitimate news footage to create convincing disinformation assets.

Google's analysts note that this still appears to be the clumsy early phase. The exploit Google intercepted had implementation mistakes that probably limited its effectiveness. But the report warns that may not stay true for long. As AI models improve and attackers refine their workflows, the quality gap between human-written and AI-generated exploits will narrow.

What the 2FA bypass means for your accounts

Here's the part that matters directly for your security. The zero-day targeted a specific web administration platform — not a consumer service. But the technique matters far beyond that single application. A 2FA bypass that works by exploiting an authentication logic flaw can, in theory, be adapted to target other systems that have similar architectural weaknesses.

Does this mean you should stop using two-factor authentication? Absolutely not. Multi-factor authentication remains one of the single most effective security measures available. Microsoft's research consistently shows that MFA blocks 99.9% of automated attacks. What this development means is that the nature of the threat is evolving — and your defences need to evolve with it.

In our analysis, the key takeaway is not that 2FA is broken — it's that authentication security needs to be treated as a layered system, not a single fix. Here's what we recommend:

• Use app-based or hardware 2FA methods — Authenticator apps (TOTP) and hardware security keys (FIDO2/WebAuthn) are more resilient against logic-based bypasses than SMS codes. Hardware keys in particular are resistant to remote bypass because they require physical presence.
• Watch for supply-chain attacks — The Canvas breach earlier this month showed how a vulnerability in an education platform's Free-for-Teacher system could cascade into a 275-million-record breach. The AI zero-day similarly exploited a vendor-side implementation error. Use a password generator to ensure unique, strong credentials, which limits the blast radius if any one service is compromised.
• Keep your software updated — This AI-discovered zero-day was patched before it could be widely exploited. The speed of the fix — Google identifying the bug, working with the vendor, and deploying a patch before the campaign launched — is a rare success story. But it only works if users actually install updates. Enable automatic updates wherever possible.
• Assume AI will find more flaws — The GTIG report makes it clear that LLMs are becoming effective at identifying high-level logic errors in authentication systems. If you run a web application, treat your authentication code as a high-value target for AI-assisted auditing. Review hard-coded exceptions, trust assumptions, and bypass mechanisms.

We ran our own small test after reading the GTIG findings. We fed an open source authentication module into a publicly available LLM and asked it to identify potential 2FA bypass points. In under a minute, it flagged three code paths that a human reviewer might have missed — including one that matched the hard-coded exception pattern Google described. It took us longer to set up the environment than it did for the AI to find the bugs. That's the speed we're dealing with now.

The bigger picture: AI versus AI in cybersecurity

Google's report is significant not just for the zero-day itself, but for what it reveals about the accelerating arms race between AI-powered attacks and AI-powered defences. Google's own security tools increasingly rely on machine learning models to detect anomalous behaviour, flag suspicious code, and predict attack patterns. The GTIG report is essentially one side of Google's AI-powered security apparatus catching the output of another AI — the criminals' model.

This dynamic — AI versus AI — is going to become the defining feature of cybersecurity over the next few years. Attackers use AI to find vulnerabilities faster and write exploits more efficiently. Defenders use AI to analyse threat intelligence, correlate signals, and deploy patches faster. The winner isn't the side with the smarter model — it's the side with the faster feedback loop.

For individual users, the practical implications are straightforward but important. The era where you could rely on a single security measure — a strong password, or two-factor authentication — is ending. The new standard is defence in depth: unique passwords for every account, hardware-backed MFA on critical accounts, regular breach monitoring, and prompt software updates. Each layer makes it harder for an attacker — even one with AI assistance — to break through.

NIST SP 800-63B guidance, updated for 2026, already points in this direction. The standard now emphasises passphrase-length passwords, breach screening during account creation, and multi-factor authentication as a baseline rather than an optional upgrade. The AI zero-day discovery reinforces why those standards matter. Password breach statistics for 2026 show that credential theft remains the most common attack vector — AI-assisted or not.

FAQs

Should I stop using two-factor authentication because of this?

No. Two-factor authentication remains one of the most effective security measures you can use. This zero-day exploited a specific code flaw in one platform — it was not a general weakness of 2FA itself. Continue using MFA, and prioritise hardware security keys and authenticator apps over SMS-based methods where possible.

Which platform had the AI-discovered zero-day?

Google's Threat Intelligence Group has not publicly named the affected platform, citing responsible disclosure practices while patches are being fully deployed. What we know: it was a popular open-source web-based administration platform with a hard-coded trust exception in its authentication flow.

How do I protect myself against AI-generated exploits?

Keep all software updated, use unique passwords for every account generated by a strong password generator, enable hardware-backed MFA on critical accounts, and monitor your credentials using Have I Been Pwned. No single measure will protect against every possible attack, but layered defences dramatically reduce the risk.

Was Gemini or Claude involved in creating the exploit?

Google confirmed that neither Gemini nor Anthropic's Mythos was used to create the zero-day exploit. The criminals used a different AI model with similar capabilities. The exploit code reportedly showed signs of LLM influence — educational docstrings, a hallucinated CVSS score, and textbook-perfect coding structure.

Is AI-powered hacking mostly a state-actor threat, or should regular users worry?

Both. Google's report identifies state-linked groups (North Korea's APT45, Chinese operators) using AI for vulnerability hunting. But the zero-day exploit Google intercepted was developed by cybercriminals — not a nation state. Google's John Hultquist said, "State actors are taking advantage of this technology but the criminal threat shouldn't be underestimated, especially given their history of broad, aggressive attacks."

How many AI-generated zero-days are out there that we don't know about?

No one knows for certain. Hultquist's comment that "for every zero-day we can trace back to AI, there are probably many more out there" suggests Google believes the visible cases represent a fraction of the total. The clumsy implementation errors in this intercepted exploit — which probably helped Google catch it — may not be present in future AI-generated exploits.

Does this mean I need to change my passwords right now?

You should always keep your passwords secure and unique. This specific zero-day did not target consumer password vaults or common authentication services. That said, now is a good time to ensure you have unique, strong passwords on every account and that MFA is enabled on all critical services. Use our free password generator to create strong, unique passwords instantly.

What are the signs that an exploit was AI-generated?

Google's analysts identified several telltale signs in this case: overly clean, textbook-style code structure with educational comments; hallucinated metadata (like CVSS scores that don't correspond to real vulnerability databases); and a polished structure that looks like it came directly from LLM training data. As AI models improve, these tells may become harder to spot.

To stay protected against evolving threats like infostealer malware and credential theft, consider a comprehensive security suite like Kaspersky Premium. It includes advanced malware protection, password monitoring, and breach alerts that help you detect compromised credentials early.