Automated Exploitation on a Massive Scale

In a concerning development for digital security, a single Chinese hacker has successfully compromised over 100 company websites in less than a month, largely relying on autonomous AI agents to perform the heavy lifting. The attacker reportedly spent roughly $5,000 in tokens to task these AI systems with the objective of gaining unauthorized access to various web platforms.

The campaign was uncovered by security researcher Snorre Fagerland of NorForge. While investigating exposed Cairn installations—an open-source AI problem-solving framework—he discovered several instances being repurposed for malicious activity, providing a rare look into an active automated cyber operation.


How the AI Agents Operated

The AI agents were tasked with bypassing website authentication to execute arbitrary code. The attackers utilized various techniques, including SQL injection, malicious file uploads, and the exploitation of unsafe stored object handling. The primary goal, according to researchers, was to secure persistent access to e-commerce and payment infrastructures to facilitate financial theft.

Fagerland noted that the agents employed a mix of models, potentially combining Claude Code with DeepSeek. The instructions provided to the AI were remarkably concise—fewer than 100 Chinese characters—yet they enabled the agents to target over 1,200 websites in under 30 days. On average, the agents spent roughly 23 hours investigating each target before either succeeding or moving on to the next, with an estimated cost of approximately $50 per successful breach.


Targeted Companies and Responses

While the initial scope of the attacks was broad, targeting online retailers across the United States, Canada, Australia, and Japan, several high-profile companies were identified, including:

  • Deere: Reported taking immediate remediation steps, stating there was «no indication that sensitive information was accessed.»
  • SharkNinja: Clarified that the breach was limited to a standalone product support site, which has since been secured.
  • Cummins: Stated that it is currently investigating the reports.

Advanced AI Capabilities in Cyberattacks

Fagerland expressed genuine concern regarding the sophistication displayed by the AI agents. He described how the systems, upon hitting a roadblock at one target, independently scanned for associated infrastructure and analyzed digital records to identify vulnerabilities. In one instance, the AI spent hours performing trial-and-error analysis until it successfully executed a malicious command on a server.

«It's not straightforward to come up with solutions like that. It's not script kiddie stuff,» Fagerland remarked regarding the complexity of the AI's adaptive tactics.

As AI-powered attacks become increasingly rapid and cost-effective, cybersecurity experts urge business owners to prioritize patching public-facing applications, restricting access to administrative panels, and monitoring web directories for suspicious or unauthorized files.