Google said on Monday that it had discovered a hacking tool capable of autonomously identifying and exploiting software vulnerabilities without human assistance, underscoring the risks posed by artificial intelligence (AI) agents.
The Google Threat Intelligence Group (GTIG) disclosed the finding in detail in its third-quarter report. According to the report, the tool was located on a pre-release server exposed to the internet. Attackers typically use such servers to store and prepare attack tools.
Google named the tool "exploit_agent." The tool connects to multiple large language models (LLMs), the AI technology behind chatbots, including leading commercial models as well as what Google calls "local" models.
The tool then monitors patch updates for popular web software, meaning updates released by software developers to fix vulnerabilities. Once a new patch is released, the tool compares the program code before and after the fix to determine where the vulnerability lies. This is similar to comparing two versions of a document to see what an editor changed, and those changes reveal where the problem is.
After that, the tool automatically writes attack code to exploit the vulnerability and executes it automatically. Google said this was the first time the company had seen attackers carry out attacks in this way.
The report said: "In the past quarter, the Google Threat Intelligence Group found for the first time that threat actors used an LLM framework to achieve full automation of vulnerability discovery and exploitation without human involvement."
The finding further reinforces growing concerns that autonomous AI agents may threaten traditional systems. Federal Reserve Governor Michael Barr said on Monday that autonomous AI agents could trigger a bank run within minutes. In September, Apollo Global Management chief economist Torsten Slok warned that banks could lose a large amount of low-cost deposits if AI assistants automatically move household funds into higher-yielding accounts.
An image in Google's report showed the tool examining the access control mechanism of version 2.19.25 of the WordPress plugin Spectra (Ultimate Addons for Gutenberg). Access control refers to the code that determines what different users are allowed to do. The report did not say whether the vulnerability check ultimately led to an actual attack.
The pre-release server also stored other scripts that could map an internal network structure after an attacker breached network defenses, which Google called "internal reconnaissance." These scripts could also conduct so-called "brute force" attacks by repeatedly trying large numbers of passwords, and plant WebShells on websites, meaning hidden remote-control programs that allow attackers to revisit infected systems later.
Google also found data stolen from victims on the same server. Related records showed that the operator of the tool may have targeted websites in the French media, entertainment and education industries. API keys are similar to passwords and allow software to access AI services.
Google's translation of the relevant Russian annotations was: "LLM API key for automated exploitation (exploit_agent pilot project)," and "Please enter the key yourself—right here, not in the chat. Both lines are required."
The finding fits a broader trend Google has previously described. Between January and August 2026, the total number of vulnerability disclosures and exploitation incidents nearly doubled. Zero-day exploits, meaning attacks targeting vulnerabilities unknown to software developers, averaged 8 per month in 2025, rose to an average of 11 per month in 2026, and surged to 22 in August.
Google also said it had found that open-source AI attack tools named CAIRN and PentAGI were used in actual network intrusion activity.
Comments