AI Agents Used to Hack More Than 100 Websites
AI Agents Used to Hack More Than 100 Websites
A cybersecurity investigation reveals how one attacker used autonomous AI agents to compromise over 100 websites for an estimated $5,000, demonstrating how artificial intelligence is changing the speed and economics of cybercrime.
One Hacker, Hundreds of Targets and an Army of AI Agents
In October 2026, cybersecurity researcher Snorre Fagerland uncovered an unusual hacking operation while investigating publicly accessible installations of Cairn, an open-source platform used to coordinate AI agents.
Fagerland discovered approximately two dozen exposed installations, some of which appeared to be connected to malicious activity. The exposure allowed researchers to observe aspects of the operation from the attacker’s own infrastructure.
According to Reuters, a Chinese-speaking attacker had been using AI agents to identify weaknesses in websites, pursue unauthorized access and carry out tasks that would traditionally require extensive human involvement.
The campaign began in late September and targeted online businesses and other websites in the United States, Canada, Australia and Japan.
More than 1,200 websites were reportedly targeted, with over 100 successfully compromised.
The central finding is not simply that a hacker used AI for assistance. The attacker reportedly delegated extended sequences of technical work to autonomous agents that could continue operating with limited human supervision.
From AI Chatbots to Autonomous Cyber Operations
Conventional AI chatbots respond to individual requests by generating text, explanations or code. Autonomous agents can take a broader role by selecting tools, executing tasks, reviewing results and deciding what to do next.
In legitimate settings, these capabilities can support software testing, website maintenance, research and administrative automation.
In the wrong hands, similar capabilities can be directed toward unauthorized activity, including searching for exposed services, identifying vulnerable applications and attempting to gain access to systems.
Automated Discovery
Agents can inspect large numbers of websites and identify potentially exposed services.
Adaptive Reasoning
AI systems can interpret technical responses and adjust their next actions.
Tool Coordination
Agent frameworks connect models to software tools and automate multistep workflows.
Extended Operation
Automated tasks can continue for hours without constant human interaction.
The distinction is important because AI agents can perform repeated technical operations at a scale that would be difficult for one individual to sustain manually.
Cairn Helped Coordinate the Autonomous Campaign
The reported operation used Cairn, an open-source platform designed to coordinate AI agents and connect them with external tools.
Investigators observed the attacker using AI models through an agent-oriented workflow that combined task planning with automated execution.
Reporting identified Claude Code and DeepSeek among the technologies associated with the operation.
The existence of legitimate AI development tools does not make those products malicious. The security concern arises when an operator deliberately directs general-purpose software toward unauthorized access and harmful activity.
The attacker reportedly supplied very brief instructions in Chinese, leaving much of the technical work to the agents.
Reuters reported that some attack instructions contained fewer than 100 Chinese characters. The agents then performed extended sequences of work, demonstrating how a relatively small human instruction can initiate a much larger automated process.
How $5,000 Became a Major Security Concern
One of the most striking aspects of the investigation is the reported cost of the campaign.
The attacker spent approximately $5,000 on AI computing credits while compromising more than 100 websites. That works out to roughly $50 per successful intrusion using a simplified estimate.
This figure is not a complete accounting of every expense involved in the operation. It does not necessarily include the attacker’s time, supporting infrastructure or other operational costs.
Nevertheless, it illustrates how automation may reduce the marginal cost of attempting attacks against large numbers of websites.
Lower Operating Costs
Automated systems can perform repetitive tasks without requiring continuous human attention.
More Targets
One operator may be able to pursue many websites over a relatively short period.
Repeated Attempts
Agents can continue evaluating outcomes and retrying tasks within their operating limits.
Geographic Reach
Internet-facing systems can be targeted regardless of national borders.
The broader risk is that attacks previously requiring substantial technical labor may become more accessible to criminals with limited resources.
Major Companies Were Among the Reported Targets
The investigation identified websites associated with well-known businesses, including Deere, SharkNinja and Cummins.
Reuters reported that some companies acknowledged security incidents while describing limited damage or restricted evidence of data compromise.
It is important to distinguish a compromised website from a breach of an organization’s entire internal network. An intrusion affecting a public-facing website does not automatically mean every connected business system was accessed.
The exact consequences vary depending on the application, its privileges, the information available and the attacker’s subsequent actions.
The investigation supports reports of more than 100 website compromises. It does not establish that all affected organizations suffered the same damage, that every website exposed sensitive customer data or that all internal corporate systems were compromised.
AI-Assisted Attacks Are Becoming a Broader Industry Concern
The NorForge investigation is part of a wider discussion about how artificial intelligence is changing offensive and defensive cybersecurity.
Security researchers have documented other campaigns involving AI-assisted exploitation, credential theft and attacks against e-commerce infrastructure.
The Cloud Security Alliance published research in September describing a separate operation involving autonomous AI tools and payment-card skimming against online retailers.
That earlier case should not be confused with the NorForge investigation. Although both involved AI-assisted cybercrime and website compromises, they were reported as distinct operations.
The growing availability of agent frameworks may make it easier to automate technical work, but AI agents are not infallible. They can misinterpret information, make mistakes, waste resources or fail against properly secured systems.
Even with these limitations, defenders face a challenge when automated attackers can evaluate large numbers of systems continuously.
What Website Owners Can Do Against Automated Attacks
Organizations do not need to deploy AI agents of their own to take meaningful precautions against automated intrusion attempts.
Fundamental website-security practices remain essential, especially when attackers can search for common weaknesses at scale.
Maintain supported versions of your content-management system, plugins, themes, server software and programming-language runtimes.
Enable multifactor authentication, use strong unique credentials and restrict privileged accounts.
Review security logs for unusual request patterns, repeated authentication failures and unexpected administrative activity.
Consider properly configured web application firewalls, rate limits and vulnerability monitoring as additional safeguards.
Maintain tested backups, protect credentials and establish a response plan for suspected website compromise.
Review installed plugins and themes, remove unused components, maintain a supported PHP version and verify that administrative accounts have only the permissions they need. Automated attackers often benefit from weaknesses that remain unpatched or unnecessarily exposed.
When Both Attackers and Defenders Use Artificial Intelligence
Autonomous AI agents are changing how security work can be performed.
For defenders, AI can assist with analyzing logs, identifying suspicious patterns, prioritizing vulnerabilities and responding to incidents.
For attackers, similar automation can accelerate reconnaissance and reduce the effort required to test large numbers of potential targets.
The result is an evolving competition between offensive automation and defensive security systems.
Organizations will increasingly need to evaluate not only whether their software is secure, but also how quickly they can detect and respond to repeated automated activity.
Strong identity controls, maintained infrastructure, careful monitoring and tested incident-response procedures remain central to that effort.
The NorForge case also illustrates an unexpected weakness on the attacker’s side: exposed agent infrastructure gave researchers visibility into the malicious operation. Poor operational security can expose attackers just as it can expose legitimate organizations.
AI Is Changing the Economics of Cybercrime
The reported compromise of more than 100 websites by a single attacker using autonomous AI agents marks a significant development in the cybersecurity landscape. According to the NorForge investigation, the attacker relied on automated tools to perform extended technical operations while spending an estimated $5,000 in AI credits. The case demonstrates how agent-based automation can reduce the labor required to pursue large numbers of targets, potentially allowing relatively small operations to generate substantial security risks. It does not mean every AI agent is dangerous or that traditional website defenses have become obsolete. Instead, it reinforces the importance of maintaining secure software, limiting privileges, monitoring suspicious activity and responding quickly to vulnerabilities. As AI becomes more capable, the advantage will increasingly belong to organizations that combine strong security fundamentals with the ability to detect and contain automated threats.
Information reflects research and reporting available through October 22, 2026. The website compromise counts and estimated costs are based on the reported investigation. The extent of impact at individual organizations may vary.
