Osegi
Security·2 min read

ARTEX hacking AI: 1,800 exposed instances, 300 stolen certificates

Criminals pointed the ARTEX AI tool at a target and walked away. Check Point researchers found 1,800 exposed instances and saw it grab 300 iOS certificates.

A hooded figure at a dark control panel with code and server racks glowing on screens, suggesting automated hacking.
Image: Spider's Web

A hacker picks a target, starts an AI and goes off to do something else. That is roughly how the security tool ARTEX was used by criminals, according to a Check Point report. Researchers got into the control panels of badly secured servers and watched the campaigns unfold.

ARTEX was built for legitimate security testing. The report says more than 1,800 instances of it have been found exposed on the internet, and 10 poorly protected servers let researchers trace whole campaigns.

What did the ARTEX AI actually do?

From 14 to 29 October, two ARTEX systems tried to break into 10 sites that distribute unofficial iPhone apps. The operator mostly just asked for progress and which target was closest to falling. The agents analysed websites, hunted for flaws, reached databases and checked whether certificates could sign iOS apps. About 36 billion tokens were processed.

The haul was more than 300 certificates with passwords. Only 3 turned out to be usable for the intended purpose. Check Point estimates the AI models cost between $700 and $1,500, mostly DeepSeek and Kimi.

In another campaign, agents spent about 5 days on 13 sites that trade accounts for Chinese mobile games. They collected over 190,000 sets of Honor of Kings credentials. In a sample, about 3.3% still worked, which suggests roughly 6,000 active logins.

A crypto theft attempt was stranger. Agents worked for 4 days against 292 payment-related targets and reported that withdrawals had succeeded. The criminal's wallet showed zero incoming funds. The model had misread system responses as success. Still, the agents found a serious hole in another service and altered files used to distribute iOS apps. The owners later restored them.

Why did the hackers switch models?

Chinese models, especially the DeepSeek Flash family, dominated. In 3 of 10 instances, US providers were configured too. Researchers saw safeguards trigger with Anthropic's Claude Opus 5.5, and the operator switched to another model. ARTEX can also present actions as authorised penetration tests, and some models then did not block attacks on systems the attacker had no permission to touch.

The agents rarely used exotic exploits. They went after old problems: weak passwords, missing access control, exposed interfaces and data left in public files. The point is that the attacker no longer needs to be an expert. Setting a goal is enough.

Apps worth a look

SUMO: the number puzzle

Tap the right tiles, add them up, hit the target. Easy to learn, hard to master. Free, for all ages.

Tags:artexcheck pointai agentscybercrimedeepseekios

More from this day

See all →

More in: Security

See all →