Check Point: ARTEX agents stole iOS signing certificates and game logins with little human input

ARTEX, an open-source AI penetration-testing platform from China, has already appeared on this site twice: in CrowdStrike's report on the China-nexus group VAPOR PANDA and in WorkOS's account of a breach of an internal sales app. A report from Check Point Research, published around November 5, adds a different view. Instead of state-linked espionage, it describes ordinary cybercriminals handing whole campaigns to AI agents and mostly just checking in on progress.

The researchers say they found 2,557 systems that host or once hosted ARTEX panels, 1,608 of them live when they looked. Ten were misconfigured and exposed their full databases, showing the prompts, models and steps behind each operation. Chinese models dominated, led by DeepSeek's Flash family. Check Point says the only guardrail refusals it saw came from an Anthropic model, after which the operator moved to more permissive Chinese models.

Certificates and game accounts

In the largest case, one operator ran two ARTEX consoles from October 14 to 29 against 10 gray-market shops that sell Apple enterprise signing certificates, using DeepSeek v4.1 Flash and Kimi K3.1 Pro. The agents exploited flaws such as admin passwords hardcoded in JavaScript, unauthenticated APIs and command injection, and came back with more than 300 certificates and passwords. Only three passed every check Check Point applied, including Apple revocation and expiry status, which would make them resellable. The run consumed about 36 billion tokens, the vast majority served from cache, which the researchers price at roughly $700 to $1,500.

A separate gaming campaign targeted the backend servers of 13 sites trading hacked mobile game accounts. Over about five days it recovered more than 190,000 Honor of Kings credential sets, one site keeping 21.5 MB of them in plain text in its web root. Sampling suggested about 3.3% were still valid, or roughly 6,000 accounts, for an estimated $180 to $360 in model costs. Another operator's attempt to drain cryptocurrency from 292 payment domains produced 230 findings but, according to the report, moved no money.

What it shows, and what it does not

Check Point's main point is that the agents worked almost autonomously once given a goal, letting operators attack at a speed and scale beyond their own skills. The figures come from leaked attacker databases and the researchers' own validation; cost estimates are based on public API prices, not observed bills. Several targets were themselves illicit services, and the report does not name victims. Check Point's advice is limited: expect high volumes of automated attacks and test and harden internet-facing systems.