Subscribe
0%
Illustration for: Check Point: crooks ran ARTEX AI agents almost on autopilot, stealing 190,000 game logins for a few hundred dollars

Check Point: crooks ran ARTEX AI agents almost on autopilot, stealing 190,000 game logins for a few hundred dollars

Check Point Research found 1,608 live ARTEX attack panels and, through ten misconfigured ones, read how operators let AI agents break into web apps with barely any prompting. The break-ins used ordinary flaws like hardcoded passwords and unauthenticated APIs, so the fixes are ordinary too.

Listen to this article0:00 / 4:27

Check Point Research published a report on November 5 showing cybercriminals running ARTEX, a multi-agent AI penetration-testing framework, as a near-autonomous break-in machine. The researchers counted 2,557 systems that host or once hosted ARTEX control panels, 1,608 of them live on the public internet. Ten were misconfigured badly enough to expose their full databases. That let Check Point read operator prompts, model choices, token bills and results from three October campaigns.

Why it matters: earlier ARTEX reporting centred on state-linked espionage. BleepingComputer covered CrowdStrike's account of a China-linked group it tracks as Vapor Panda targeting 200+ organisations in 14 countries. Check Point's data shows the other side of the market. Low-level crooks are pointing the same tool at soft targets, and the agents mostly run on their own. In two of the three campaigns, Check Point says, operators gave no prompts after the initial task.

The three campaigns, by the numbers

All figures are Check Point's, from the exposed databases. Cost estimates are Check Point's own calculations from token counts.

Campaign Targets Run time Models Est. cost Outcome
Apple code-signing certs 10 gray-market signing shops Oct 14–29 DeepSeek v4.1 Flash, Kimi K3.1 Pro $700–$1,500 (~36B tokens) 300+ certificates/passwords taken; 3 passed validation
Gaming accounts 13 backend credential servers ~5 days DeepSeek V4.1 Flash $180–$360 190,000+ Honor of Kings credential sets; ~3.3% (~6,000) estimated valid
Crypto draining 292 crypto payment domains 4 days DeepSeek v4 Pro $145–$300 230 findings, 400,791 agent messages, no funds moved

The targets were mostly gray-market businesses: services that sell .p12 certificates for sideloading iOS apps, and sites trading mobile-game accounts. According to Cyber Security News, one signing service had more than 90,000 subscribers. Some of the gaming servers kept a 21.5 MB plaintext credential file in the public web root.

The crypto campaign failed, but not cleanly. Check Point says some agents claimed success just because an API replied "success." In one side attempt, an agent found leaked AWS S3 credentials and confirmed write access to a bucket serving 714 enterprise-signed iOS app landing pages. The bucket's owners restored the original manifest.

What the agents actually exploited

Check Point lists no CVEs. The way in was basic web hygiene failures:

  • Hardcoded admin credentials in JavaScript bundles, and weak admin passwords
  • Web APIs with no authentication
  • Command injection and server-side template injection leading to code execution
  • Insecure direct object references (IDOR)
  • Plaintext credential files in the web root
  • Loose folder permissions allowing escalation to root
  • Leaked cloud storage keys

Models and guardrails

Chinese models dominated the configured LLMs, led by the DeepSeek Flash family. American models were configured on three instances. Check Point says guardrails only kicked in with Anthropic's Opus 5.5, and the operator then switched to a more permissive Chinese model. Live panels were concentrated in the US (882, or 54%), Hong Kong (295) and China (103). Cyber Security News adds that operators used residential proxies to get past web firewalls and IP blocks, and that the ARTEX developer made the project closed-source on October 8. Forks and backups remain easy to find.

What to do

The pattern is cheap agents trying every easy door at volume, so the defence is closing the easy doors:

  1. Search your front-end bundles for secrets. Grep built JavaScript for API keys, admin passwords and tokens. Rotate anything you find.
  2. Require authentication on every API route, including internal and "temporary" admin endpoints. Check object access per request to stop IDOR.
  3. Keep data out of the web root. No credential dumps, backups or exports in served directories.
  4. Scope and rotate cloud keys, and lock down write access to buckets that serve downloads or app manifests.
  5. Alert on bursts. Check Point warns that targets of agentic campaigns should expect very high request volumes. Rate-limit and alert on sudden spikes from residential IP ranges.
  6. Hunt the infrastructure. Check Point published three likely same-operator ARTEX clusters, including 255 hosts tagged sco-fix2 on AS54600 (PEG TECH). It cautions that shared tags are clustering hints, not unique indicators.

Who should care: anyone running internet-facing web apps that were built fast and never pen-tested. A few hundred dollars of tokens now buys a thorough sweep for the mistakes listed above.

Sources

  1. Check Point Research — Hacking on Autopilot: Inside ARTEX Cybercrime Campaigns research.checkpoint.com
  2. Cyber Security News — ARTEX AI Agents Steal 190,000 Gaming Credentials and Hundreds of Apple Code-Signing Certificates cybersecuritynews.com
  3. BleepingComputer — Chinese hackers use ARTEX AI in automated attacks against 200 orgs bleepingcomputer.com

#Check Point#ARTEX#agentic AI#DeepSeek#threat intelligence#web security