In July 2026, Chinese hackers launched autonomous AI agents to launch cyberattacks on retailers around the world. The bots were tasked with deploying digital skimmers to steal payment data. Since then, the automated system has launched hundreds of attacks, compromising dozens of organizations and stealing at least 600,000 bank card records. The campaign is still active.
In July 2026, Chinese hackers launched autonomous AI agents to launch cyberattacks on retailers around the world. The bots were tasked with deploying digital skimmers to steal payment data. Since then, the automated system has launched hundreds of attacks, compromising dozens of organizations and stealing at least 600,000 bank card records. The campaign is still active.
This was reported by cybersecurity researchers from Gambit, who managed to intercept the attackers' server and recreate the course of this large-scale operation. In just five days, from September 10 to 15, AI agents carried out 105 waves of attacks and hacked 27 organizations to varying degrees, TechRadar reports .
The cybercriminals' victims included a Fortune 500 hospitality company, a major U.S. airline, a large privately held industrial goods distributor, and an American fashion retailer. One AI tool used website ranking analysis services to find potential targets, focusing on resources with proprietary software.
What struck researchers most was not the sheer number of victims, but the efficiency and cheapness of the cyberattack. According to Gambit, Chinese hackers are behind these actions, using three artificial intelligence frameworks. These systems are capable of almost completely autonomously carrying out the entire attack chain, affecting about 10 companies per day.
Over four weeks, the attackers spent only about $7,000, and the total cost of the operation for the entire time did not exceed $18,000. On average, hacking one company cost hackers $25.46 (from $3.13 for the simplest targets to $79.31 for the most secure).
“If bots were able to gain access, it usually took less than a day, and in many cases just a few hours. We also found instructions in the AI instructions for deleting data and performing cleanup procedures, which in some cases led to disruption of the operations of the affected companies,” Gambit noted.
To manage the campaign, the hackers used a complex of three AI tools:
Hermes is an open-source autonomous AI agent that acted as an operator console. On the server, the developers found a Chinese system profile “SOUL — Red Team Operator,” which contained 121 skills (78 of which were offensive). Hermes ran on the Anthropic Opus 4.6 model (after newer models began blocking malicious requests). The human operator only provided brief instructions in Chinese, running 260 sessions with a total of 1,951 prompts — just a few requests per target.
Strix is an open source pentesting tool.
Cairn was a standalone intrusion engine that ran on DeepSeek v4.1 Flash. The tool was given a target domain and a task (e.g., gaining root access or a web shell) and worked autonomously for hours until it achieved its goal.
Gambit researchers noted the persistence, patience, and creativity of AI agents, which are difficult for a human hacker to achieve. They called on businesses to urgently change their approaches to security: “Organizations must adapt to a reality where attacks occur much faster and are more global in nature. To do this, it is necessary to implement defense systems that can counteract artificial intelligence at its own speed.”
Most of the affected companies have already been informed of the leak and the implemented skimmers have been removed from their sites. However, the investigation and elimination of the consequences of this campaign are ongoing.
