CyberXero Deploys Swarm of 51 AI Agents Against Ukrainian Critical Infrastructure

Researchers at SOCRadar uncovered the CyberXero campaign after discovering an openly accessible HTTP server at 46.21.250.135 that exposed the group's full workspace, logs, tools, and stolen data. The operation ran two parallel tracks using the same infrastructure: one mass campaign against WordPress and e-commerce sites worldwide, and a targeted strike on Ukraine's energy sector and critical infrastructure. Attackers created 51 specialized AI agents inside the Claude Code environment to automate reconnaissance, vulnerability identification, exploit chaining, and data exfiltration. They also integrated the PentAGI framework with Cobalt Strike for payload generation and attack automation while bypassing model refusals by claiming authorized penetration tests. The group relied on known vulnerabilities such as CVE-2026-4815 in WooCommerce rather than zero-days, achieving rapid results including 4,708 hosts scanned and 32 webshells deployed in 61 seconds. The campaign resulted in the theft of 628,000 records, primarily from a Kharkiv heating supplier, containing names, addresses, account numbers, and energy consumption details.

ekurakTranslated from Polish

Read at source

Researchers at SOCRadar have detailed a sophisticated cyber campaign by the group CyberXero that leveraged a swarm of 51 AI agents to target Ukrainian critical infrastructure and global e-commerce platforms.

The discovery occurred during routine monitoring when analysts found an unprotected HTTP server at 46.21.250.135. The server exposed the attackers' working directory, including reconnaissance tools, scripts, logs, and exfiltrated data, representing a significant OPSEC failure.

Parallel Operations

CyberXero conducted two simultaneous operations using shared infrastructure. One campaign focused on mass exploitation of WordPress sites and e-commerce platforms worldwide. The second targeted Ukraine's energy sector, specifically a heating supplier in Kharkiv.

Attackers avoided zero-day exploits and instead chained known vulnerabilities. Key vectors included SQL Injection via CVE-2026-4815 in WooCommerce, remote code execution flaws in popular WordPress plugins, and the wp2shell tool for deploying webshells. They also performed large-scale reconnaissance using custom orchestrators combined with nuclei and Python scripts.

AI Agents in Action

The most notable element was the use of 51 specialized AI agents created in the Claude Code environment. Each agent handled specific tasks such as scanning, vulnerability detection, and exploitation. The group further integrated the PentAGI framework with Cobalt Strike to generate payloads and automate core attack phases.

When AI agents refused tasks, operators bypassed restrictions by claiming they were conducting authorized penetration tests on their own systems. Persistent refusals were overcome by opening new sessions to reset context.

In just 61 seconds the operation scanned 4,708 targets, confirmed access to 429 WordPress admin panels, and deployed 32 webshells. The campaign ultimately yielded over 628,000 records belonging to Ukrainian citizens, including names, addresses, bank account numbers, and detailed energy consumption data.

While primarily a financially motivated Initial Access Broker, the focus on Ukrainian critical infrastructure suggests possible state alignment or approval.

Meridian digestThe week's most consequential security stories, distilled into one briefing.See the digest

More in AI & Model Risk

AI Evolution Drives New Approaches to Vulnerability Management at Internet Week 2026
Architecture in the Age of Probability: Why Legacy Blueprints Fail in AI-Integrated Banking Systems
AI Hackers vs AI Defenders: Arms Race Accelerates Cyber Espionage in 2026