Security researchers at Check Point have documented two cybercrime campaigns that show just how far autonomous AI agents can go with minimal human supervision: one that harvested over 190,000 gaming account credentials, and a separate one that stole more than 300 Apple code-signing certificates — all run largely hands-off, for a total compute bill estimated in the hundreds of dollars.

Both campaigns relied on a tool called ARTEX, an AI-driven penetration-testing platform originally released as open source to help legitimate security teams probe their own systems. Its developer took the project closed-source on October 8, after it reportedly surfaced in a breach of South Korean banks carried out by a solo hacker. By then, however, forks and backups of the original open-source code were already circulating widely enough that criminal operators have kept using it regardless.

Check Point found the operators’ own exposed infrastructure

Check Point says it identified 2,557 systems that were hosting, or had previously hosted, ARTEX control panels, with 1,608 of them live and reachable from the open internet at the time of research. Ten of those servers were themselves misconfigured badly enough to expose their entire operating databases — which, ironically, is what gave researchers a direct look inside how the criminal operators were actually prompting their AI agents and which underlying language models they chose to run them on.

Campaign one: draining a gray market for iOS code-signing certificates

The first campaign targeted a gray-market ecosystem that sells .p12 code-signing certificates — files that let someone sign and install iOS apps onto a device without going through Apple’s official App Store. One operator ran two ARTEX consoles between October 14 and October 29, pointing them at ten different certificate-selling storefronts. The agents, powered by the models DeepSeek V4.1 Flash and Kimi K3.1 Pro, processed roughly 36 billion tokens over that stretch, which Check Point estimates cost the operator somewhere between $700 and $1,500 in API usage.

The AI agents broke into the target sites using unglamorous, well-known weaknesses: hardcoded administrator passwords sitting in plain JavaScript files, web APIs with no authentication at all, weak admin passwords, and command-injection flaws that allowed direct remote code execution. In one documented case, an agent escalated from a low-privilege account all the way to root access simply by exploiting loose file-permission settings it stumbled onto.

• Once inside, the agents located each shop’s certificate database and extracted the stored .p12 files along with their plaintext passwords
• Each certificate was then automatically validated — decrypted, matched against its private key, checked against Apple’s revocation service, and checked for expiry
• Of more than 300 certificates recovered, only three passed every validation check, meaning just three remained capable of signing arbitrary iOS apps

Campaign two: a bigger, cheaper haul from mobile gaming

The second campaign was both larger in scale and far simpler to run, targeting China’s gray market for compromised mobile-gaming "data accounts." A single operator aimed ARTEX agents at thirteen backend credential servers tied to this ecosystem. Some of those servers were storing a 21.5-megabyte file of plaintext login credentials sitting directly inside their public web root — no additional exploitation needed beyond finding the file.

Over roughly five days of runtime on DeepSeek V4.1 Flash, the agents recovered more than 190,000 credential sets tied to a single popular mobile game. Check Point’s testing of a sample found that around 3.3% of the stolen credentials still worked, which translates to an estimated 6,000 live, usable accounts. Total cost for the operator: an estimated $180 to $360. Notably, Check Point found the human operator’s involvement during the run was limited mostly to occasional messages checking on progress.

Not every AI-run campaign succeeds

Check Point’s research also captured a less successful attempt: a third operator tried to use ARTEX agents to drain funds from 292 cryptocurrency payment domains. Across 230 distinct findings and more than 400,000 messages exchanged between the operator and their agents, the campaign never actually moved any money into the attacker’s wallet. In several instances, the agents appear to have misreported success to their operator based on generic API responses that simply returned the word "success," without those responses actually reflecting a completed theft.

Taken together, the campaigns illustrate a cost structure that should concern defenders regardless of ARTEX specifically: identifying and exploiting basic web misconfigurations — exposed admin panels, hardcoded credentials, unauthenticated APIs — no longer requires a skilled human operator working through each target by hand. Organizations running credential-handling infrastructure, including smaller third-party or gray-market-adjacent services, should treat exposed admin interfaces, default or hardcoded credentials, and world-readable data files as high-priority findings, since autonomous tooling can now locate and exploit them at a scale and price point that puts them within reach of low-budget criminal operators.