A new cybersecurity report is raising serious questions about the extent to which artificial intelligence can already be used for automated attacks. According to an investigation published by researcher Snorre Fagerland, a Chinese-speaking hacker used AI agents to attempt intrusions into more than 1,200 websites, leaving the tools to handle nearly all of the technical work.
The campaign began in late September
According to the report, the attacks began on September 25 and mainly targeted e-commerce websites in the United States, Canada, Australia and Japan. Among the affected entities, the report mentions major international companies such as Deere, SharkNinja and Cummins.
The central claim of the report is that the operation was not directed step by step by a human, but instead relied on autonomous AI agents that searched for weaknesses, tested intrusion paths and adapted their strategy during the attack on their own.
Minimal instructions, prolonged action by AI
According to the researcher, the commands given to the agents were very short, fewer than 100 characters in Chinese. Despite this, the AI systems carried out lengthy searches for vulnerabilities and continued for hours without direct human intervention.
The report estimates that a successful intrusion required an average of about 23 hours of uninterrupted research and testing. By the end of last week, according to the same source, more than 1,200 attacks had been attempted against different websites.
A single case shows the level of automation
One of the episodes Fagerland considers most significant involves a large company, where the AI agents initially did not find vulnerabilities in the public-facing application. Instead of stopping, they analyzed data and digital traces to identify infrastructure connected to the system.
Then, according to the report, an internet-exposed control panel and a vulnerability in the way it processed data were discovered. After several hours of testing, the agents managed to exploit it to make the server execute malicious commands.
Low cost, greater concern
Fagerland estimates that the entire campaign consumed about $5,000 in artificial intelligence tokens, or roughly $50 for each successful intrusion. According to him, a combination of DeepSeek and Claude Code technologies appears to have been used in this case.
Reuters reported that it was not possible to determine whether the attacks resulted in unauthorized theft of data. Of the companies mentioned, Deere said it had taken corrective measures and had no indication that sensitive data had been accessed, SharkNinja said a separate support website had been affected, while Cummins said it is verifying the reporting of a possible attack on a website operated by a service provider.
The case is being seen as a strong signal of how autonomous AI agents may significantly reduce the cost and human effort involved in cyberattacks.
That is exactly what makes the warning more serious: not necessarily because every claim in the report has been fully verified, but because the technical threshold for sophisticated attacks appears to be falling rapidly.