Google finds AI agents autonomously turning patches into exploits
Researchers warn that AI agents have reached an important turning point. They can now find security weaknesses and use them to launch attacks automatically.

According to a report released Monday by Google Threat Intelligence Group (GTIG), attackers are using large language models (LLMs) and groups of AI agents to find and attack security weaknesses without human help. Researchers called this a "critical turning point" in the use of AI for cyberattacks. Their findings show a growing gap between attacks that happen at machine speed and defenses that depend on people. They warn that automated attackers could soon move faster than company security teams can respond.
GTIG said it had seen attackers use LLM-based systems to complete every step of finding and exploiting security weaknesses without a person involved. Between July and September 2026, researchers found an exposed test server running a system called exploit_agent. This system uses several AI agents and works with different advanced and locally run LLMs. It constantly compares patched and unpatched versions of popular web applications to look for possible security weaknesses. It then creates attack code and runs it automatically, without human help.
"By chaining these capabilities together, threat actors can dramatically reduce the window between patch disclosure and weaponized exploitation," GTIG researchers wrote.
Researchers also found scripts designed to help attackers after they break into a system. These scripts can search inside networks, try many passwords and install webshells. Information from affected systems showed that the operator targeted French websites in the media, entertainment and education industries.
Earlier versions of Google's quarterly AI threat reports had described only some parts of attack processes being automated.
Company patching schedules are under pressure
GTIG researchers also pointed to attacks in August 2026 as evidence that attackers are putting fully automated systems for finding and exploiting security weaknesses into use. These attacks targeted weaknesses in PaperCut print management servers. According to researchers at GreyNoise, the attackers appear to have used LLMs to create exploits and target hundreds of organizations in a large campaign.
The report adds to concerns about the faster use of n-day vulnerabilities. These are known security weaknesses that can remain unfixed in company systems for months or even years. Experts warn that as tools for comparing software patches and creating exploits improve, defenders will have much less time to fix vulnerable systems.
"Traditional patching cycles are becoming obsolete, and some fixes will need to happen in hours rather than days or weeks," said Jason Elrod, chief information security officer (CISO) at MultiCare Health System.
Open-source tools that let several AI agents work together to carry out attacks are now widely available. These include Artex, Cairn, CyberStrikeAI and PentAGI. GTIG warned that attackers will keep finding new ways to use these tools. Many security leaders say they also need tools with similar abilities to defend against attacks that happen at machine speed.
"We have to bring AI-enabled defense to bear in order to combat AI-enabled offense," said Bob Leek, chief information officer (CIO) of Clark County, Nevada. He added that using automation is one of the county's main cybersecurity goals for the coming year.
AI attackers focus on speed rather than hiding — for now
GTIG said that today's AI attack systems focus on speed rather than staying hidden. They create large amounts of network activity and leave traces in code, making them easier to detect. But researchers expect this to change as AI tools improve.
Attackers who care most about staying hidden and keeping access to systems, such as cyberespionage groups, are likely to avoid AI tools that leave many traces behind. However, researchers warned that many cybercriminals who want to make money will choose speed over secrecy and use automated tools. They said this is likely because reaching more targets could make up for a lower chance of success in each attack.
This strongly suggests that the pressure on defenders will keep growing in the coming months.
"AI cuts attack timelines from days to minutes, so detection-to-containment time has to shrink the same way," said Neil Wyler, vice president of defensive services at cybersecurity services provider Coalfire and senior network operations lead at Black Hat.
Alissa Irei is a TechTarget senior reporter covering cybersecurity news.
