Google discovers dangerous AI hacking tool: cyberattacks without human intervention
Google discovers an AI tool that can independently exploit security vulnerabilities. Crypto companies are also being attacked, while the Federal Reserve warns of financial risks.
Google has discovered a new form of cybercrime in which artificial intelligence (AI) can independently find and exploit security vulnerabilities. The technology makes it possible to carry out attacks without hackers having to control every step. According to Google, this is a significant development, especially now that banks and crypto companies are increasingly being targeted by sophisticated cyberattacks.
On Monday, November 16, researchers from the Google Threat Intelligence Group (GTIG) announced that they had discovered a noteworthy program on a publicly accessible server belonging to cybercriminals.
The software, calledexploit_agentuses multiple AI models to detect vulnerabilities in computer systems. The program monitors security updates for popular software. Once a developer fixes a bug, the AI compares the old and new program code.
This allows the system to identify where the original security vulnerability was located. It then independently writes and tests code to exploit that vulnerability.
According to Google, this is the first time that researchers have discovered such a fully automated system among attackers. On the same server, they also found stolen data and tools that hackers can use to guess passwords and maintain access to compromised systems.
The data found indicate interest in French websites in the media, entertainment and education sectors. Google has not determined that every feature of the AI tool was actually used on victims.
Google also discovered new attack methods used by a suspected North Korean hacker group, which the company calls MIDNIGHT NEPTUNE. This group allegedly attacked crypto companies in North America and Europe using AI.
Among other things, the criminals created fake recruiter profiles, complete with AI-generated profile photos. They then lured employees to counterfeit meeting websites, where victims were encouraged to run malicious software.
Once inside a company, the attackers modified the settings of the AI programming tool Claude Code. This allowed a backdoor to be automatically installed when a developer opened the program.
The researchers also observed that the hackers were creating access keys with names that referenced Microsoft Copilot, presumably to avoid attracting attention.
The findings come in the wake of the major Binance hack in October, in which, according to U.S. authorities, approximately $6.3 billion was stolen. However, Google has not determined thatexploit_agentwas used in that attack.
The warnings are not limited to cyberattacks. Michael Barr, a board member of the U.S. Federal Reserve, warned on Monday that autonomous AI systems can also pose risks to banks.
According to Barr, AI assistants could independently move large amounts of savings in the event of financial turmoil. This could significantly accelerate a bank run, in which customers withdraw their money en masse.
According to him, major banks have already encountered AI systems that attempted to carry out transactions on customer accounts without adequate verification of consent. Barr therefore advocates stricter identity checks and clear restrictions on what AI is allowed to do independently.
Meanwhile, Google is observing that cybercriminals are exploiting security vulnerabilities at an ever-increasing pace. In August, 22 so-called zero-day vulnerabilities were attacked, compared to an average of eight per month in 2025.
With such vulnerabilities, the software developer has not yet had the opportunity to fix the security issue.