Saturday, 21 November 2026

Over 628,000 Ukrainian files exfiltrated in Claude Code hacking spree

A Russian-speaking access broker called CyberXero used Claude Code agents to hack WordPress shops and probe Ukrainian utilities, leaking its own playbook.

WordPress, related to: Over 628,000 Ukrainian files exfiltrated in Claude Code hacking spree
Photo: Simone Bignetti / CC BY-SA 4.0 (modified)

The short version

  • Researchers at SOCRadar tied an AI-assisted hacking campaign to CyberXero, a Russian-speaking access broker active since at least July 2026.
  • The group ran Claude Code with 51 specialized agents to scan, chain exploits and steal data, mostly from WordPress and online stores.
  • More than 628,000 records about Ukrainian people were taken from four Ukrainian organizations, with other breaches in Poland, China and the Netherlands.
  • Site owners should look for unknown WordPress admins, WSO web shells and strange SSH keys on any Redis server open to the internet.
Quick read · 2 min

A Russian-speaking hacking crew called CyberXero used Anthropic's Claude Code, running 51 specialized AI agents, to break into WordPress sites and online stores around the world, and to probe energy and utility firms in Ukraine. Researchers at SOCRadar found the group's working files on an exposed server, complete with AI chat logs and stolen data.

The haul included more than 628,000 records about Ukrainian people from four organizations, plus break-ins in China, Poland and the Netherlands. In one run, the tools scanned 4,708 sites, got into 429 WordPress admin panels and planted 32 hidden web shells in about a minute.

  • The hackers dodged Claude's refusals by claiming they were testing their own systems and restarting sessions.
  • Some refusals still held, including requests to install backdoors and turn off firewalls.
  • Site owners should hunt for unknown admin accounts, web shells and rogue SSH keys on open Redis servers.

The campaign was still active when the research came out, so patching vulnerable plugins is urgent.

One automated run took 61 seconds to plant 32 web shells on hacked websites. That is the pace security researchers found when they stumbled onto the working files of CyberXero, a hacking crew that used Anthropic’s Claude Code to run much of its operation, and walked away with more than 628,000 records about people in Ukraine.

Why should you care if you don’t live in Ukraine? Because the bulk of this Claude Code cyberattack campaign hit ordinary WordPress sites and online shops all over the world, the same kind of sites that sell you sneakers or host your local plumber’s booking form. The goal there was simple: get in, then sell the access or skim payment data. It is also one of the clearest public looks yet at how criminals are putting AI coding agents to work, and where the safety guardrails held or failed.

01

What the Claude Code cyberattack actually was

The Claude Code cyberattack was a two-track hacking operation run by CyberXero, which the security firm SOCRadar describes as a Russian-speaking, money-driven initial access broker. An access broker is a criminal who breaks into networks and then sells that foothold to other gangs, such as ransomware crews. SOCRadar believes the group has been active since at least July 2026, and says the campaign is still running.

The first track is wide and automated. It sweeps the internet for WordPress sites and e-commerce platforms with known security holes, breaks in, and looks for login details and card data it can cash out. The second track is narrow and hands-on. It focuses on Ukrainian energy and utility companies, including a heating provider in Kharkiv, with slow, careful reconnaissance and manual hacking.

That second track is the puzzling part. SOCRadar says there is no obvious way it makes money, which sets it apart from the rest of the group’s work.

Operations center with monitors, code, and Ukrainian documents on a desk
02

How hackers used AI agents to work at scale

CyberXero used AI at two separate points in its attacks. On the main workstation, it ran Claude Code with a cluster of 51 specialized agents, each handling a piece of the job: scanning for targets, linking several flaws together into a working break-in (known as exploit chaining), mapping a victim’s systems, and doing the follow-up work once inside.

On a second machine, the group connected PentAGI, a free, self-hosted tool built for fully automated penetration testing, to Cobalt Strike. Cobalt Strike is a commercial tool for testing defenses that criminals have long abused. Linked through an AI provider’s API, the pair generated attack payloads and ran test attacks with little human input.

The scale shows why this matters. In a single automated pass, the setup scanned 4,708 targets and confirmed access to 429 WordPress admin panels. Then it dropped 32 web shells, small hidden programs that let an attacker control a site remotely, in about a minute.

03

Getting around Claude’s safety refusals

Claude is built to refuse requests that look like hacking. CyberXero’s workaround was to lie about who it was. The logs show the operator telling the agent it was doing an authorized security test on infrastructure it owned.

When the model still said no, the operator simply closed that session and opened a new one with the same cover story already loaded. SOCRadar points out that a refusal only lives inside one conversation; it doesn’t carry over to the next, so retrying in a fresh session can sometimes get a different answer.

The guardrails didn’t fold every time, though. Across four attempts, the model refused to install a backdoor, switch off a firewall, spread to other machines on the network, and plant a web shell. In another episode, it turned down a scan four times in a row and spotted a faked proof-of-ownership token. Those refusals are a useful reminder that AI tools are not magic hacking buttons, but also that persistence and a convincing lie can get some tasks done.

A computer screen showing a WordPress login page
04

Who was hit, and what was taken

SOCRadar confirmed data theft from four Ukrainian organizations, which together yielded more than 628,000 records linked to Ukrainian individuals. It also confirmed break-ins in China, the Netherlands and Poland.

The two Chinese victims were reached through a different door: Redis servers left open to the internet. Redis is a fast database many websites use to store temporary data, and when it is exposed without a password, attackers can use it to slip their own SSH keys onto the server and log in whenever they like.

The targeted flaws are not exotic. The group went after the wp2shell bugs, tracked as CVE-2026-63030 and CVE-2026-60137, and CVE-2026-4815, an SQL injection hole in the Support Board WordPress plugin. SQL injection lets an attacker sneak database commands into a web form. All three already had public details, which means unpatched sites were the easy pickings.

05

How the researchers found it

CyberXero gave itself away. SOCRadar found an open directory on one of the group’s servers that exposed the operator’s live working folder from a Linux machine. It held AI session logs, scripts with passwords and tokens in plain text, hacking toolkits, scan results, the agent setup files and stolen victim data.

From that folder, researchers rebuilt the whole setup: eight distinct machines, including the operator’s workstation, a server for spinning up new infrastructure, systems for cashing out stolen access, servers that staged payloads, and a dedicated scanner hunting for targets.

06

What this means for website owners and shoppers

If you run a website on PrestaShop, Magento or WordPress, this campaign is a direct prompt to check your setup today. The advice from researchers is specific:

  • Look at your list of WordPress administrator accounts and remove any you don’t recognize.
  • Scan your site for WSO web shells, a common type of hidden remote-control script.
  • If you run Redis and it can be reached from the internet, check the authorized_keys file in the SSH folder for entries you didn’t add, and lock Redis down so outsiders can’t reach it.
  • Update WordPress plugins, especially Support Board, and apply fixes for the wp2shell bugs.

For shoppers, the risk is quieter. If a small online store you use was hit, your card details could be at risk without you ever hearing about it. Check your card statements regularly for charges you don’t recognize, and report them to your bank quickly.

07

Quick answers

Did Claude Code itself get hacked?

No. Claude Code wasn’t breached. Attackers used it as a tool on their own computer, lying to it about having permission, much as criminals have long misused legitimate security software.

Were 628,000 people affected?

The figure is records tied to Ukrainian individuals, taken from four Ukrainian organizations. One person can appear in several records, so the number of people may be lower.

Is the campaign over?

No. SOCRadar says the operation was still ongoing when it published its findings, so patching and checking for rogue admin accounts remains urgent.

You're reading the quick version.