Hackers are using artificial intelligence to study software security fixes and turn them into working attacks, according to a new report from Google’s threat intelligence team. Researchers said criminals have moved past typing single prompts into chatbots. They now run groups of AI agents that carry out attacks on their own.
The report was published Monday. It said attackers are using AI to shrink the time between when a software flaw is fixed and when it is exploited.
How Patches Become Targets
When a company releases a security patch, it shows what was broken in the old code. Attackers compare the new version against the old one to find the flaw. This process is called differential analysis.
The report said large language models can now automate much of this work. Hackers can feed the models source code from different versions, including preview and pre-release builds. They can also add release notes, security research and existing proof-of-concept code.
Because the source material is limited, AI tools can work through it quickly. This lets attackers focus on “n-day” flaws, which are known bugs that have been patched but not yet fixed on every system.
Researchers pointed to attacks on PaperCut NG/MF printing software in September as one example. In that case, attackers used AI to weaponize newly disclosed flaws.
Attack Tools Built on AI Agents
The report named several open source tools being misused, including ARTEX, CAIRN, CyberStrikeAI and PentAGI. These were built for penetration testing, which is when security teams test their own systems for weak spots. Attackers are using them to automate steps during live break-ins.
Other tools named include ZeroHollow, a service used for exploit development and for routing requests between AI models. Agent Zero, an open source framework that can run code and build new tools on its own, was also listed.
Researchers said these tools let attacks run at machine speed. Attackers can link different attack paths together and scale up their operations, even when the basic methods stay the same.
Several groups have tested these frameworks, the report said. They include criminals seeking money and suspected spy groups linked to China.
Researchers also found attackers stealing access to AI systems themselves. In attacks known as LLMjacking, hackers take AI login details and cloud resources from hacked company networks.
That stolen access is then resold at lower prices through underground services such as APIMart and JetNight. These services are known as LLM mixers.
The report also described a Chinese-language phishing platform called Jieyou. The platform added an AI chatbot that lets customers build and change phishing pages using plain language commands.
This means users do not need technical skills to set up fake login pages or change bait messages. The report described AI as a “modular force multiplier” for phishing-as-a-service sellers.
Researchers said open source frameworks have made it easier for less skilled attackers to get started. These tools help them add new exploits, scale attacks on their own and run more convincing phishing campaigns.
The report said the result is a shorter attack timeline. The gap between a patch going public and active exploitation is getting smaller.