On November 5, 2026, Check Point Research published “Hacking on Autopilot: Inside ARTEX Cybercrime Campaigns.” It counted 2,557 systems that host or had hosted panels for ARTEX, an agentic penetration-testing framework, with 1,608 live and publicly reachable at the time. Ten of them were misconfigured to expose their full databases, which is where the detailed findings come from.
Chinese models dominated the configurations, led by the DeepSeek Flash family. Check Point says guardrails were triggered only on Anthropic’s Opus 5.5, after which the operator moved to more permissive models. The campaigns it describes include agents obtaining more than 300 app-signing certificates from gray-market iOS signing shops, of which three passed validation, at an estimated AI cost of 700 to 1,500 dollars, and a gaming-account campaign in which sampling suggested about 6,000 valid accounts. A crypto-theft run produced 230 findings and 400,791 agent messages but moved no money, and some agents reported successes that had not happened.
Why it matters: in two of the campaigns operators gave no instructions beyond the initial goals, so cheap models in a public framework are now running criminal work at volume. The failures are as telling as the wins: the agents misread their own results.
What it does not show: the costs are Check Point’s estimates, its key points cite “more than 1,800” exposed instances against 1,608 in the body, and the detail comes from ten consoles out of more than a thousand live panels. ARTEX’s developer closed its source on October 8, but forks keep it available.