Google detects AI zero-day
- Google’s threat-intelligence team said on May 11 it caught a criminal group preparing a mass attack with a zero-day exploit it believes AI helped create. - The exploit hit a popular open-source admin tool, bypassed two-factor authentication after login, and showed LLM fingerprints like tutorial-style docstrings and a fake CVSS score. - That matters because the AI race in hacking is no longer theoretical — Google says real attackers are already using it to shrink discovery time.
A zero-day is the nightmare version of a software bug. Nobody outside the attacker knows it exists, so there is no patch, no signature, and basically no warning. Google says it has now seen something worse — a real-world zero-day exploit that appears to have been discovered and developed with help from AI, then lined up for a broad criminal campaign. Google says it warned the affected vendor before the attackers could go wide, and the vendor patched the flaw. ### What actually happened? Google Threat Intelligence Group published the finding on May 11, 2026, in a broader report on how attackers are using generative AI. The headline claim is narrow but important: this is the first time GTIG says it has identified a threat actor using a zero-day exploit that it believes was developed with AI. The group says the actor intended a mass-exploitation event, and that Google’s “counter discovery” likely stopped it before launch. (cloud.google.com) ### What was the bug? The vulnerability was in a widely used open-source web-based system administration tool. The exploit let an attacker bypass two-factor authentication once the attacker already had valid credentials. That means this was not a magical one-click internet takeover — the criminal still needed usernames and passwords first — but it turned stolen logins into much more reliable access. The flaw came from a logic mistake: one part of the software trusted something that another part was supposed to verify. (cloud.google.com) ### Why does “AI-developed” matter here? Because the hard part of zero-days is usually not just finding a bug. It is turning that bug into a working exploit quickly enough to use before defenders catch up. Google says the code itself carried signs of LLM involvement — educational docstrings, a hallucinated CVSS severity score, and very clean, textbook-style Python. None of that proves a model did everything. But it does suggest AI was not just used to draft phishing emails or summarize research — it was in the exploit-development loop. (helpnetsecurity.com) ### Did Google say which model did it? Not publicly. Google says it does not believe Gemini was the model involved. Politico reported that Google also concluded Anthropic’s Claude Mythos was likely not the source. So the important point is not “which lab’s model did this.” The point is that at least one capable model appears to have been good enough for a criminal to move from bug hunting to weaponization. (helpnetsecurity.com) ### Is this the first AI cyberattack? No — but it is the first case Google says involved an AI-developed zero-day exploit in the wild. Attackers have already been using AI for faster phishing, malware tweaks, research, and obfuscation. Google’s same report says Russia-linked actors used AI-generated decoy code in malware, while Chinese and North Korean actors have shown strong interest in using AI for vulnerability work and operational support. (politico.com) ### Why is mass exploitation the scary part? Because cybercriminals do not need a perfect exploit if they can use it fast against lots of targets. A zero-day aimed at broad deployment can become a smash-and-grab event before defenders even know what to look for. AI changes the timing. It can shorten the gap between “there’s a weird bug here” and “here is working code to abuse it.” That compressed window is the real story. (cloud.google.com) ### So what is Google doing about it? Google is making the obvious argument: if AI helps attackers find and weaponize bugs, defenders need AI on their side too. The company points to Big Sleep for vulnerability discovery and CodeMender for automated fixes, while also tightening abuse controls around Gemini accounts and model access. That does not solve the whole problem, but it shows where the industry is heading — machine speed on both sides. (cloud.google.com) ### What’s the bottom line? The big shift is not that hackers used AI at all. They have been doing that for a while. The shift is that Google says AI has now crossed into one of the most dangerous parts of offensive cyber work — building a zero-day exploit that could have been used at scale. That means the old comfort line, that AI was mostly helping with low-level criminal chores, is getting harder to believe. (cloud.google.com) (blog.google)
