Exposed AI pentesting panels let one hacker breach 100+ sites

NorForge, a security research firm, published a report on 19 October saying it found more than 20 control panels for Cairn, an open-source AI-driven penetration-testing engine, sitting open on the public internet with no authentication. Searching for them with Shodan and Censys, researchers logged over 1,100 attack projects across the panels, most from a Chinese-language cluster targeting e-commerce and payment infrastructure in the US, Japan and Australia.

Cairn chains AI agents through reconnaissance with Playwright browser automation, then exploitation, hunting for SQL injection, insecure file uploads, deserialisation bugs and authentication bypasses. NorForge says the average project took just under 23 hours and cost around $5 in AI tokens, with longer runs passing $100. One documented chain combined a memcached session store with unsafe PHP deserialisation and a Zend Framework gadget chain to plant a webshell. Running the exposed panel data through the Trufflehog secrets scanner turned up hundreds of secrets, 11% of them valid credentials.

Why it matters: the vulnerability classes here are the familiar ones, SQL injection, bad upload handling, insecure deserialisation, not some novel AI exploit. What has changed is the cost of finding and chaining them: a few dollars in tokens now buys hours of autonomous probing that used to need a skilled manual tester. That is a cheaper adversary for exactly the bug classes code review and dependency scanning exist to catch.

The caveat: NorForge is working with Reuters to notify victims and does not name the breached companies in its own report. The panels were exposed by operator misconfiguration, not a flaw in Cairn’s own code.