Research · 3 min read
Google says some intruders are now letting agents find and use flaws
The company’s November 16 threat update says open attack frameworks are shortening the time between a patch and an exploit. It points to the same ARTEX tooling the FBI described last week.

Google’s threat intelligence group says the groups it watches have moved past typing questions into a chatbot. On November 16, 2026 it published an update arguing that some of them now run multi-agent tools that look for flaws and try them, with less of a person in the middle. The post is Agentic Threat Activity Crosses a Significant Threshold.
What Google says has changed
The update, written with Mandiant, says financially motivated groups and espionage groups are both using generative tools to compress the time an intrusion takes. The practices it names include autonomous frameworks, pipelines that turn a published patch into an exploit, and the resale of stolen access to language models, which it calls LLMJacking.
The frameworks it lists as already abused in real intrusions include ARTEX, CAIRN, CyberStrikeAI and PentAGI. We wrote about the FBI’s November 15 advisory on ARTEX in a separate post. Google’s point is that ARTEX is one tool in a set, not a one-off.
On finding flaws, Google says some actors use models to compare a program before and after a patch, then try the difference against servers on the internet, without a person stepping through each stage. It describes one exposed system, which it calls exploit_agent, that was aimed at French media, entertainment and education sites. Google says it found that system because the operator left a staging server open, with stolen data and webshells on it. The company does not, in the summary, say how many sites were actually entered.
It also says the flood of supposed proofs of concept after PHP2Shell, many of them machine-written, made defenders’ work slower: each sample had to be checked. We covered the Check Point attacks that used those PHP flaws in another post.
Google’s forecast, as BleepingComputer reports the same update, is that within three months some groups will be able to run unsophisticated attacks with little or no human interaction, and that more fully automatic attacks are likely within six months. A forecast is not a count of attacks that have already happened.
What you can do
- Patch internet-facing software when the vendor ships a fix, and do not wait for a newsletter. Google’s claim is that the gap between a patch and an exploit attempt is getting shorter.
- Treat a public proof of concept as untrusted code, especially in the first days after a bug is announced. Google says many of the PHP2Shell samples were noise.
- If you run the tools Google names, assume other people do too. The FBI’s advice from the ARTEX advisory still applies: close unused admin pages, limit cloud administration, and read sign-in logs for bursts of automated attempts.
- If you are not running servers, this update does not ask you to change a password. It is about the people who operate the services you use.
How Frame helps
Frame can also look up your verified emails in Have I Been Pwned’s breach database. This report is about attack tooling, not a new list of addresses. Frame doesn’t scan your servers.
Related coverage
- The FBI says attackers are using the ARTEX framework to speed up break-ins: the advisory Google’s update points at.
- Project Zero is shortening the time vendors get before a bug is published: the disclosure clock Google’s own research team shortened three days earlier.
- Check Point says attackers used PHP2Shell, then wiped some VMware hosts: the PHP flaws Google uses as an example of a noisy proof-of-concept flood.
Sources
- Google Threat Intelligence Group, Agentic Threat Activity Crosses a Significant Threshold, November 16, 2026
- BleepingComputer, Google warns that fully-automated AI attacks are coming soon, November 16, 2026
More stories
New posts, by email.
About once a month. Plain-language privacy notes, nothing else. I never share your address.


