Cybercriminals are increasingly using artificial intelligence (AI) to carry out cyberattacks, even when they have limited technical skills.
Researchers have found that four previously unreported hacking groups used AI tools to conduct attacks and bypass safeguards designed to prevent these models from helping with harmful activities, according to CrowdStrike .
The findings , published on Wednesday, showed that the groups operated in Chinese, Spanish, English, Indonesian, and Russian. Researchers from cybersecurity firm CrowdStrike studied the AI systems used by these groups in attacks carried out between May and August.
The researchers found that the first two groups had developed sophisticated systems, known as AI harnesses, which are software tools that guide AI models in completing tasks. The other two groups had less technical expertise but still managed to compromise multiple targets.
This allowed researchers to examine their activities and identify the AI models they used. These included different versions of Anthropic’s Claude and the open-weight DeepSeek models, which allow users to access their underlying model weights. The groups also designed their campaigns to operate automatically for relatively long periods, reducing the need for constant human involvement. Adam Meyers, CrowdStrike’s senior vice president of counter adversary work, explained that attackers can use different AI models for different tasks. Since newer models can be expensive, choosing the right model for each activity helps reduce costs and avoid using more powerful tools than necessary. The researchers observed that the hackers used both open-weight and closed-weight models. In almost all cases, they relied on older versions rather than the most advanced models. Meyers noted that legitimate AI users also frequently choose older models to manage rising costs. The attackers also used multiple AI models to exploit differences in their safety protections. These safeguards are designed to stop AI systems from assisting with harmful activities, but they do not work identically across all models, including those developed by the same company. Instead, the groups took advantage of AI’s speed and ability to perform tasks at a large scale, targeting common security problems that could have been prevented but were left unresolved. The researchers found that all four groups used methods to get around AI safeguards. One common approach was to convince AI models that their activities were legitimate and authorised, such as work being done for penetration testing or security evaluations. Some groups placed files in an AI agent’s working directory to make it appear that their tasks had been approved. An AI agent is a system that can carry out tasks with a degree of independence. The hackers also used traditional jailbreaking techniques. These involved creating prompts and stories that presented their activities as harmless and authorised. They also tried to address questions or concerns that an AI model might raise about the requested tasks. The more technically advanced groups used carefully planned scenarios to convince AI models that their activities were legitimate. The less sophisticated groups sometimes used more direct and aggressive prompts to pressure the models into accepting their claims. The findings show how cybercriminals with different levels of technical expertise are using AI to automate attacks and work around safeguards, even without developing new hacking methods.
All four groups left their AI infrastructure openly accessible on web servers, according to WIRED . Meyers said the attacks did not depend on new hacking techniques or previously unknown security weaknesses.
