TLDR
- Google’s threat team says attackers are using AI agents to find and exploit software flaws with no human in the loop.
- Hackers use AI to compare software updates and quickly build attacks for newly fixed bugs.
- A framework called exploit_agent was found on an attacker’s server, linked to attacks on French websites.
- Criminals are also stealing AI access from companies and reselling it at a discount.
- Google expects simple autonomous attacks within one to three months.
Google’s Threat Intelligence Group published its Q4 2026 AI Threat Tracker on November 16. The report says criminals have moved past typing prompts into chatbots. They are now running autonomous AI tools that carry out several steps of an attack on their own.
Hackers are now using LLMs to continuously scan software updates, spot newly fixed bugs, and build attacks before teams can apply the fix — all without human involvement. That’s according to the latest report from the Google Threat Intelligence Group.
We're helping defenders…
— News from Google (@NewsFromGoogle) November 16, 2026
Researchers worked with Mandiant on the findings. The report names tools such as ARTEX, CAIRN, CyberStrikeAI and PentAGI. These frameworks are built to automate parts of a break-in while it is happening.
How Patches Become Attack Tools
When a company fixes a security bug, it releases an update. Attackers can compare the old and new versions of the code to see what changed. That comparison points them to the flaw.
WHAT HAPPENS WHEN HACKERS NO LONGER NEED HUMANS?
Google has documented the first observed fully autonomous AI-powered cyberattack, where AI independently discovered vulnerabilities and attempted exploitation.
Think about the implications for crypto exchanges, DeFi protocols and… pic.twitter.com/U8Hbj67IWZ
— Crypto Patel (@CryptoPatel) November 17, 2026
Security researchers have done this for years. Google says AI now lets attackers do it faster and with less skill. Google pointed to flaws in PaperCut NG/MF software this September as one example.
The report says attackers have an easier job because they work from a limited set of material. That includes code versions, release notes, security research and proof-of-concept code. AI tools can read all of it and write an exploit.
Google says this shortens the time between a fix being released and an attack starting. Organizations that are slow to install updates face more risk as a result. Some of the attackers tied to these tools are financially motivated, and others are suspected spies aligned with China.
What Investigators Found on the Server
Investigators found a compromised staging server that belonged to an attacker. It held a multi-agent framework called exploit_agent. The framework connects to several large language models on its own.
It then studies the code of popular web applications before and after patches. Google says it runs the full loop by itself and keeps going without a person watching. It finds weaknesses and writes working exploits.
Google calls this the first fully automated discovery and use of software flaws by malicious actors that it has seen. In its Q3 2026 report, the team had found AI-assisted workflows but no fully autonomous systems. Earlier in 2026, AI mostly helped with tasks like harvesting credentials.
The campaign linked to the server mostly targeted French websites. The victims were in media, entertainment and education. Stolen data was found on the server, which shows the tool was used in real attacks.
Investigators also found scripts for use after a break-in, including brute-force tools and web management systems. Google says this points to a well-organized operation. It also says automated attacks can move fast and leave few traces.
The report also describes a trend called LLMjacking. Attackers steal AI credentials and cloud resources from hacked company systems. They then resell the access through underground services such as APIMart and JetNight at lower prices.
Open-source tools are lowering the skill needed to launch attacks. Google named ZeroHollow, a service used for exploit development, and Agent Zero, an open-source framework that can run code on its own. It also named a Chinese-language phishing platform called Jieyou.
Jieyou includes a chatbot that lets customers set up fake login pages with plain language commands. Google says less skilled users can change phishing lures without editing any code.
Google says defenders will need automated tools of their own to keep up. The company is offering Google AI Threat Defense, which uses Gemini and other models to find and fix flaws before attackers can use them.
Google expects simple autonomous attacks to appear within one to three months. It expects more advanced versions within three to six months.







