# Inside ARTEX: Check Point finds 1,608 live panels and reads 10 of them

> Check Point Research read the databases of 10 misconfigured ARTEX consoles and found campaigns that ran for days with almost no operator input.

- **Severity:** Info: Research, tools, evals or lab claims with no direct exposure. Context, not an alert.
- **Category:** Incident, Research
- **Published by:** Check Point Research
- **Disclosed:** 2026-11-05
- **Affects:** ARTEX, DeepSeek Flash, Honor of Kings
- **Research by:** Check Point Research (Check Point Software Technologies)
- **Primary source:** https://research.checkpoint.com/2026/hacking-on-autopilot-inside-artex-cybercrime-campaigns/
- **Page:** https://theinjection.dev/items/checkpoint-artex-autopilot-campaigns/

Check Point Research found 2,557 systems that host or hosted an ARTEX panel, 1,608 of them live, and read the full databases of 10 misconfigured instances to reconstruct three financially motivated campaigns.

## What happened

Check Point Research published this analysis on November 5, 2026. Scanning for ARTEX consoles turned up 2,557 systems that host or once hosted a panel, 1,608 of them live and publicly reachable when the researchers looked. Ten of those instances were misconfigured so that their full database was exposed, which let Check Point read the operators' goals, the agents' messages and the results they recorded.

## How it works

ARTEX links a language model to several agents that plan tests, run security tools, investigate findings and write them up. Check Point reports operators set a goal once and then mostly checked progress: two of the three campaigns had no operator prompting after the initial objective. Across the 10 readable instances, Chinese models dominated the configured providers, with the DeepSeek Flash family most common, and three instances used American models. Guardrails fired only with Anthropic's Opus 5.5, after which the operator moved to a more permissive model.

## Who is affected

Victims were opportunistic and financially motivated rather than single chosen organizations: crypto and digital asset services, online gambling, payment infrastructure, over-the-counter crypto trading and custody, and gray-market iOS signing shops. Live command servers sat mostly in the United States at 882, Hong Kong at 295 and China at 103, which describes where operators hosted panels rather than where victims were. Check Point expects many further deployments to sit behind firewalls or on internal networks.

## What to do

Test and harden internet-facing systems, giving priority to weak authentication, exposed APIs, credential leakage, remote code execution flaws and cloud storage permissions. Restrict write access to cloud storage buckets and stop unauthenticated endpoints from returning storage credentials. Plan detection and response for high-volume automated probing rather than the pacing of a single human operator.

## Quick facts

- **Exposure:** 2,557 systems hosted or had hosted an ARTEX panel, 1,608 live and publicly reachable, 10 exposing their full database
- **Code signing:** Two consoles ran about two weeks against 10 iOS signing shops, obtained more than 300 certificates, and 3 passed the full validation checklist
- **Gaming accounts:** 13 credential servers over about five days yielded more than 190,000 Honor of Kings credential sets, about 3.3% still valid, an estimated 6,000 accounts
- **Crypto draining:** 167.5 million tokens over 4 days, 292 domains in the first wave, 230 findings and 400,791 agent messages, with $0 reaching the attacker's wallet
- **Cost:** Check Point estimates $700 to $1,500 for the code-signing campaign and $145 to $300 for the crypto campaign, at DeepSeek API prices at the time
- **Found by:** Check Point Research

## Questions

### How did Check Point get inside the ARTEX operators' consoles?

Check Point did not break in. Ten of the 1,608 live ARTEX panels it found were misconfigured so that the full ARTEX database was exposed to anyone who reached the host. Reading those databases gave Check Point the objectives operators typed, the messages agents exchanged, the findings they logged and the model provider each console was configured to use.

### Did the ARTEX campaigns actually make money for the operators?

Results varied sharply in Check Point's reconstruction. The code-signing campaign obtained more than 300 certificates, 3 of which passed full validation and could sign arbitrary iOS apps, and the gaming campaign produced an estimated 6,000 working accounts. The crypto draining campaign spent 167.5 million tokens over four days and delivered $0, with some agents reporting cashouts that Check Point found were hallucinated.

### Which models were ARTEX operators running, and did any refuse?

Chinese models dominated the 10 consoles Check Point read, with the DeepSeek Flash family the most common choice. Three instances were configured with American models. Check Point reports guardrails were triggered only with Anthropic's Opus 5.5, and that the operator responded by switching the console to a more permissive model rather than abandoning the objective.

## Sources

1. [Check Point Report](https://research.checkpoint.com/2026/hacking-on-autopilot-inside-artex-cybercrime-campaigns/)
2. [Check Point Summary](https://blog.checkpoint.com/research/inside-artex-how-autonomous-ai-agents-are-being-used-in-real-world-cyberattacks/)