Autonomous AI frameworks accelerate intrusions, according to Google and the FBI
Google and the FBI describe a rise in autonomous AI frameworks in cyberattacks. ARTEX, CAIRN, CyberStrikeAI, and PentAGI are used to automate intrusion, phishing, and exploitation of vulnerabilities.
Google states that cybercriminals are now using autonomous AI frameworks such as ARTEX, CAIRN, CyberStrikeAI, and PentAGI to automate several stages of an attack. The FBI reports having tracked campaigns linked to ARTEX targeting organizations worldwide, with privileged access obtained in less than three hours in some cases. This development matters because it reduces the time needed to breach a network, increases the number of attackers able to act, and makes intrusions easier to industrialize.
Faster, less human attacks, therefore harder to stop
The change is not only technical. It alters the pace of malicious operations. According to Google Threat Intelligence Group, these agent tools allow tasks to be distributed among several components: reconnaissance, scanning, exploitation, then maintaining access. In other words, attackers no longer need to do everything manually. The machine handles part of the work, and it does so quickly.
On November 13, the FBI issued a flash alert on ARTEX, described as being used since at least August 2026 against organizations worldwide. Targeted sectors include financial services, government, technology, defense, semiconductors, energy, healthcare, industry, telecoms, and legal services. In some cases, attackers moved from the initial target to elevated privileges in less than three hours. For security teams, the countdown is therefore even more ruthless than before. Cybercrime already thrives on bad days; it has just learned to time them.
ARTEX, increasingly widespread tooling
ARTEX reportedly appeared on GitHub in July 2026, uploaded by a user named Autumn-27. The project, written in Go, includes a web console in Next.js and connects to external models such as Claude, GPT, DeepSeek, GLM, Grok, and Kimi. According to the FBI, Chinese models are most often observed among attackers, especially DeepSeek Flash, considered attractive for its capabilities, lower cost, and lack of visible limits during penetration tests.
The original repository was reportedly deleted in October 2026, but the tool continues to circulate as community forks, notably ScopeWeaver, and via a commercial version called ARTEX PRO. This distribution matters as much as the code itself: the simpler a tool is to reuse, the faster it can descend the criminal chain.
LLMs also help find exploitable vulnerabilities faster
Google observes another rising use: language models are mobilized to analyze software patches and speed up exploit creation. Attackers compare versions, read release notes, examine security articles, and cross-reference these elements with already published proof-of-concept exploits. The goal is clear: identify what the patch has not fully closed, then turn that remainder into an exploitable vulnerability.
This method reduces the gap between vulnerability disclosure and exploitation. Google cites the PaperCut NG/MF vulnerabilities as a recent example. Here, AI does not magically discover bugs, but it saves time for those who already know where to look. And in this field, time is often the most expensive resource.
Automated phishing and resale of stolen AI access
Phishing operations are also gaining momentum with AI. Google mentions Jieyou, a Chinese phishing-as-a-service platform that integrates an LLM chatbot. Users can give it natural language instructions to create phishing pages and adjust templates without advanced technical skills. This lowers the entry barrier for less experienced affiliates.
Added to this is another phenomenon: LLMjacking, meaning the theft of LLM credentials and cloud resources in compromised environments, then resale via clandestine proxy services like APIMart and JetNight, also called LLM mixers. In this scheme, AI is no longer just an attack tool; it also becomes stolen merchandise. The black market never lacks imagination when it comes to monetizing what does not belong to it.
Why this shift changes the game
The common point among these uses is simple: AI does not replace the attacker, it equips them better. Groups motivated by money as well as espionage actors can now automate reconnaissance, exploitation, phishing, and access resale. As a result, campaigns are more numerous, faster, and accessible to less expert profiles.
For companies, the issue goes beyond the mere AI trend. It affects detection speed, patch management, credential monitoring, and cloud service protection. At this pace, security teams will need to monitor not only the humans behind the attack but also the agents working for them.
Key points
- Google cites ARTEX, CAIRN, CyberStrikeAI, and PentAGI.
- The FBI mentions access in less than three hours.
- ARTEX has targeted global organizations since August 2026.
- Jieyou automates phishing in natural language.
- LLMjacking: theft then resale of AI and cloud access.
- DeepSeek Flash is often used by attackers.
By the numbers
- 3 hours — time observed to obtain elevated privileges, according to the FBI.
- 67,000 people — personal data stolen during attacks in South Korea, according to the FBI.
- 2,250 downstream clients — exposure linked to an attack targeting ConnectWise Cloud Backup.
- November 13 — date of the FBI flash alert mentioned in the sources.
- August 2026 — presumed start of ARTEX-related attacks, according to the FBI.