Check Point finds 1,608 live ARTEX panels exposed on the public internet, alongside 949 systems that previously hosted the AI penetration-testing framework. Ten misconfigured consoles exposed their full databases, giving researchers an unusually detailed view of how operators directed autonomous agents during real campaigns.

 

The November 5 findings materially extend earlier reports linking ARTEX to attacks on South Korean financial companies. They do not establish that every reachable server was criminal: legitimate security teams can deploy the same open-source software, and one operator may control many panels.

 

The operator-side evidence adds four important findings:

  • Researchers identified 2,557 current or former ARTEX hosts.
  • Ten consoles exposed prompts, agent messages and campaign records.
  • Several campaigns ran with little input after the initial objective.
  • Results ranged from stolen credentials to costly model hallucinations.

 

Related Coverage

 

Check Point Finds 1,608 Live ARTEX Panels

Check Point Research identified 2,557 systems that hosted or had hosted an ARTEX command panel. Of those, 1,608 were live and publicly reachable when the researchers measured them, while additional deployments may have remained behind firewalls or inside private networks.

 

ARTEX is an LLM-driven framework originally designed for authorized penetration testing. An operator states a security goal, and multiple agents break it into tasks, run reconnaissance and testing tools, record findings and pursue promising attack paths through a dashboard.

 

The project does not contain its own model. Operators connect an external model, so capability and safeguards vary by deployment. The original developer removed the public repository after reports of misuse, but independent forks and backups kept the software available.

 

Check Point found that 882 of the live servers were hosted in the United States, 295 in Hong Kong and 103 in China. Those locations describe infrastructure, not the nationality or physical location of an operator, and they should not be treated as attribution evidence.

 

Exposed Databases Reveal Nearly Autonomous Campaigns

Ten ARTEX servers were configured so poorly that their databases were exposed. Check Point used those records to examine operator goals, model choices, agent conversations and the technical steps taken during campaigns, moving beyond victim-side telemetry that showed ARTEX artifacts without revealing how operators managed the tool.

 

DeepSeek Flash models were the most common configuration in this small sample. Three consoles used American models, and Check Point reported seeing a guardrail intervene only with Anthropic's Opus 5.5. The operator then changed models. Ten exposed consoles are too few to establish market-wide model use or refusal rates.

 

Across the observed campaigns, human involvement often concentrated on setting the objective and checking progress. Agents performed scanning, vulnerability testing, exploitation and follow-on work through real tools. That persistence let them test mundane weaknesses repeatedly, including hardcoded passwords, missing API authentication, command injection and insecure object references.

 

The findings show why autonomy can matter even when the underlying vulnerabilities are familiar. Software can keep examining endpoints, JavaScript bundles, credentials and cloud permissions without the time constraints that limit a human operator, increasing the chance that several small weaknesses will be joined into a consequential attack path.

 

ARTEX Campaigns Produced Uneven Results

One operator aimed ARTEX agents at grey-market iOS signing services. The agents obtained more than 300 certificates and passwords, then validated them against key, revocation and profile checks. Check Point said three certificates passed every test and could sign arbitrary iOS applications.

 

A second operator targeted 292 cryptocurrency-related domains. Agents found serious weaknesses, including database exposure and writable cloud storage, but the attempted theft produced no blockchain transactions. Some agents incorrectly treated successful API responses as proof of a payout, while the attacker's wallet remained at zero.

 

The failed campaign processed 167.5 million tokens over four days at an estimated model cost of $145 to $300. Its 400,791 agent messages illustrate that high activity does not guarantee a successful outcome: autonomous systems can pursue false leads, misread responses and report success without independent verification.

 

A third campaign targeted gaming-account infrastructure. Agents recovered more than 190,000 credential records, and a sample suggested roughly 3.3% remained valid, implying about 6,000 working accounts. Check Point found that the agents continued after the initial task without further operator prompts.

 

CERT-EU Urges Faster Exposure and Patch Checks

CERT-EU separately warned that it had observed threat actors using ARTEX and assessed that public availability would make automated campaigns faster and more scalable. It also said activity affecting its constituency was highly likely to have involved the framework.

 

The agency's immediate guidance focuses on internet-facing systems: reduce unnecessary exposure, isolate management panels, patch urgent vulnerabilities and monitor privileged accounts. Its longer-term recommendations include multi-factor authentication, endpoint detection, web application firewalls and logging that can reveal automated scanning or repeated tool signatures.

 

For defenders, the new evidence changes the planning assumption. A weak endpoint may be tested persistently by agents that can move from discovery to exploitation without waiting for a person to review every step. Asset inventories, emergency patching and credential controls therefore need to operate against machine-paced probing.

 

The report does not show that ARTEX has made expert operators unnecessary or that every campaign succeeds. It does show that open agent frameworks can lower the cost of sustained opportunistic attacks, while their own exposed consoles can also leak operational data. Both attackers and defenders remain vulnerable to poor configuration.

 

Feature image: Sailors and civilian staff in the U.S. Fleet Cyber Command operations center at Fort Meade, photographed by Petty Officer 1st Class Samuel Souvannason. Used as an illustrative cybersecurity image; cropped and resized from the public-domain U.S. Navy photograph.