> ## Content Index
> Fetch the complete content index at: https://www.russianhackers.co/llms.txt
> Use this file to discover other available public pages before exploring further.

# AI Found the Flaw, Google Got There First: A Mass Exploit Foiled
- URL: https://www.russianhackers.co/google-ai-zero-day-mass-exploit-foiled/
- Published: 2026-07-15T15:51:42.000Z
- Updated: 2026-07-15T16:31:20.000Z
- Description: Google says it caught, with high confidence, the first known case of a criminal group using AI to find and weaponize an unknown vulnerability. They planned to deploy it at scale. It was likely stopped in time.
- Author: Martin
- Tags: AI, 0/1 day

In May 2026, Google's Threat Intelligence Group (GTIG) said it had disrupted an attempt by hackers to use artificial intelligence to prepare what the company called a mass vulnerability exploitation operation. According to GTIG, a criminal group used an AI model to find and weaponize a [zero-day](https://www.russianhackers.co/microsoft-confirms-rogueplanet-defender-0day/), a flaw unknown even to the software's developer.

## What happened

The AI-discovered vulnerability was a semantic logic error in a widely used administration tool, one that let an attacker bypass two-factor authentication. Google assessed that the group intended to deploy it in a mass attack, but that the company's proactive counter-discovery likely prevented it. Google did not name the group, and it stressed that it does not believe its own Gemini model was used. It is the first known case of a zero-day being discovered and prepared for use with the help of AI.

## Malware that thinks

The vulnerability is not an isolated case. In its report, GTIG describes a whole class of malware that, for the first time, calls on language models while it runs. One example, PROMPTFLUX, is a VBScript dropper that reaches out to the Gemini API and asks, on the fly, for fresh ways to disguise itself and slip past antivirus tools. It even has a module with a telling name, "Thinking Robot," which periodically queries the model for new evasion code.

Another example is closer to our beat. GTIG links the Russian state-backed group APT28 (also known as FROZENLAKE) to a piece of malware called PROMPTSTEAL, used against Ukraine. It is a data stealer that reaches an LLM through the Hugging Face API and asks it to generate commands during the attack itself, rather than carrying them ready-made.

## States are already trying

According to Google, generative AI, Gemini included, is being taken up by state-backed groups from North Korea, Iran and China, which apply it across every stage of their operations. North Korean crews, the report says, used AI for cryptocurrency theft, translation and deepfake creation.

## What it means

The point is not that AI has learned to hack on its own. It is that AI speeds up and cheapens work that used to demand rare skill. Finding an unknown flaw, bypassing a defense, generating malicious code on the move, AI does all of it faster than a human and without tiring. One more detail stands out: the attack was stopped not because the AI slipped, but because defenders went looking for the threat ahead of time. More and more, the other side will be a machine, and finding it will take the same, a machine against a machine.

*Based on research by Google's Threat Intelligence Group.*