Article
Hacker issued commands in Chinese, reportedly used DeepSeek to breach more than e-commerce websites, yet left their own backend unsecured
Beating AI Flash report: A hacker using Chinese-language commands delegated nearly the entire intrusion process to AI. Since September 25, the system has successively targeted e-commerce websites in the United States, Japan, Australia, and other countries, making more than 1200 attempts in total. Attack logs indicate that more than websites were breached. Reuters' verification found that affected companies included U.S. agricultural machinery giant John Deere and well-known home appliance brand SharkNinja.
The hacker first used an automated browser to screen e-commerce websites that could process orders normally and had a certain level of traffic, then handed the targets to the open-source AI framework Cairn. It can orchestrate multiple Agent simultaneously, continuously seek paths to targets, and place discovered clues on a shared task board so subsequent tasks can continue the work. The hacker needed fewer than 100 Chinese characters to specify the attack target and method, and AI would independently scan for and test vulnerabilities, switching to another approach after a failure.
On one occasion,AI could not find a vulnerability on a major company's public website, so it followed network records to trace related servers and unexpectedly discovered an externally accessible administrative panel. It then spent several hours studying how the panel processed data and, after repeated attempts, ultimately reported successfully causing the server to execute malicious commands. Security firm NorForge also independently verified a web shell left by AI on another supplier's website, confirming that these attacks were not merely successes claimed in logs.
Researchers found that the executor for attack tasks was named `claudecode_deepseek-v4.1-flash` and inferred that the hacker used Claude Code to call DeepSeek V4.1 Flash. The task success rate for the overall operation was estimated at approximately 10%, with successful intrusions taking less than 23 hours on average. Based on task scale and model pricing, total model invocation costs were approximately 5000 USD, with each website breached for only approximately 50 USD.
More ironically, the hacker's own Cairn control backend lacked adequate access protection. NorForge found 14 backends suspected of belonging to the same Chinese-language attack cluster, enabling it to view the hacker's instructions,AI execution records, and some attack results. Researchers also identified other attackers using Cairn, some of whom specifically targeted credit cards and cryptocurrencies.