The idea of a cyberattack carried out end-to-end by artificial intelligence was until recently a matter of foresight. An investigation of Check Point Researchthe research arm of the Israeli security publisher, published on November 5, 2026shows that it has become a reality. Cybercriminals have entrusted AI agents with entire hacking campaigns, from spotting targets to exploiting vulnerabilities, simply setting the initial objective and asking, from time to time, how the work was progressing.
ARTEX, a misused defense tool
At the heart of the investigation is ARTEXa Chinese penetration testing platform, that is to say a tool that simulates attacks to help an organization find its own vulnerabilities. It coordinates multiple agents based on language models, which plan tests, launch security tools, and report their results in a dashboard. Its intended use is limited to systems that one owns or is authorized to test.
Distributed as open source, ARTEX has been hijacked. According to Check Point, its developer decided to make the project closed on October 8, 2026 after its use against South Korean banks, but copies and variants remain available online. Other similar platforms exist, such as PentAGI or CyberStrike.
The researchers identified 2,557 servers on the internet hosting or having hosted an ARTEX console, including 1,608 still active at the time of the study. More than half are in the United States, ahead of Hong Kong and China. Above all, ten of these installations were incorrectly configured and exposed their complete database: instructions given by operators, models used and agent activity log. A rare window into the daily work of cybercriminals.
For its part, the company Censys has identified, according to TechCrunch, more than 800 ARTEX installations accessible without a password, specifying that not all of them are malicious and that many are used for authorized tests.
Three campaigns scrutinized
Check Point details several campaigns carried out in October 2026:
| Campaign | Duration | Result according to Check Point | Estimated cost of models |
|---|---|---|---|
| Theft of iOS signing certificates | Two weeks, from October 14 to 29 | More than 300 certificates recovered, three of which passed all checks | $700 to $1,500 |
| Attempted theft of cryptocurrencies | About four days | No funds stolen, but a breach exposing online storage space | $145 to $300 |
| Theft of video game accounts | About five days | More than 190,000 identifiers, of which approximately 3.3% are still valid | $180 to $360 |
The first campaign targeted gray market stores that offer to sign iOS apps outside of the App Store. A valid signing certificate allows you to install an application on an iPhone without going through the official Apple store. The agents stole more than 300 certificates and their passwords, then sorted them themselves: three passed all the checks, which makes them, according to Check Point, tools that can be used to sign any application and likely to be resold. The study does not specify whether Apple revoked these certificates.
The third campaign targeted sites linked to the resale of accounts for the mobile game Honor of Kings. One of them stored his identifiers in plain language in a publicly accessible file. More than 190,000 identifiers have been recovered; out of a sample, 3.3% were still functioning, or around 6,000 accounts according to the researchers’ estimate.
Almost autonomous agents
The most striking observation concerns the role of humans. According to Check Point, most attacks were entirely autonomous, apart from the definition of the initial objective. The operators’ instructions most often consisted of asking for progress or designating new targets. The agents were responsible for spotting, searching for vulnerabilities, exploiting them, verifying the results and writing reports.
This observation contrasts with that of Microsoft, which estimated at the end of October that the operations of North Korean hackers using AI remained controlled by humans. Autonomy, however, has its limits: in the campaign targeting cryptocurrencies, certain successes announced by the agents were based on simple misinterpreted technical responses, and Check Point describes part of these results as hallucinations.
Inexpensive and minimally restricted models
The first campaign consumed approximately 36 billion tokensthe units of text processed by the models, the majority consisting of data reused from a buffer memory, billed less expensively. At the official price of DeepSeek, Check Point estimates its cost between 700 and 1,500 dollars.
Operators mainly used Chinese models, DeepSeek v4.1 Flash and Kimi K3.1 Pro. According to the researchers, American models were configured on only three installations, and only Claude Opus 5.5, from Anthropic, triggered guardrails, which led the operator to turn to more permissive models. A reminder that protections built into models matter, but an attacker can always choose another.
Commonplace flaws, exploited on a large scale
The study reassures on one point: the agents have not invented new techniques. They exploited known weaknesses, but with a patience and scale that no human team could afford at this price. Among them, Check Point notes:
- of the administrative passwords written in the code websites;
- accessible programming interfaces without authentication;
- weak admin credentials;
- identifier files stored in plain text on a web server;
- online storage spaces open for writing while serving public content.
According to TechCrunch, many ARTEX-related attacks also took advantage of weak or known passwords and the lack of multi-factor authentication.
Check Point warns that organizations targeted by this type of campaign should expect a high volume of attacks and recommends testing and hardening systems exposed on the internet. Concretely, this involves known but often neglected measures:
- activate themulti-factor authentication on all administration access;
- never leave a password or key in the code sent to browsers or in a public folder;
- verify that each programming interface requires authentication;
- restrict write rights to online storage spaces;
- treat signing certificates and their passwords as high-value secrets, and revoke them if in doubt.
A dramatic drop in the cost of the attack
The major lesson of the survey is economic. Campaigns that would have mobilized a team for weeks now cost a few hundred dollars in calculation. Defenders, especially small structures, will have to deal with attackers capable of relentlessly testing every gap left open. The answer involves less new tools than basic cybersecurity applied without exception, and defense agents capable of finding these vulnerabilities before the attackers.

