CrowdStrike details agentic AI campaigns by four threat groups using older Claude and DeepSeek models between May and August.
The analysis highlights how actors bypassed model safeguards through prompt techniques and runtime patches, assisted by openly accessible server infrastructure.
The four groups communicated in Chinese, Spanish and English, Indonesian and Russian. The AI systems used by these groups were left open online, which allowed researchers to study their actions. The attacks mainly used older versions of the Anthropic and Deepseek models.
How they bypassed security measures
To circumvent rules prohibiting the use of artificial intelligence for illegal purposes, the perpetrators convinced the models that their actions were authorized, such as being under security control. The methods they used included adding files to working folders and specially crafted manipulation texts. The Chinese-speaking group even attempted to remove internal security instructions from the system they were using.
Objectives and attacks per team
The group, which operated in Spanish and English, targeted healthcare organizations in Argentina and managed to compromise at least 15 organizations. From at least one victim, the attackers stole login credentials and medical records, while also creating an autonomous attack system using different AI roles.
READ ALSO
Hackers from Belarus hid in Russian health organization for two years
The Chinese group targeted government agencies in Brazil and Colombia, as well as private companies in sectors such as biotechnology, health, food, trade, technology and tourism. It mainly used specific models from Anthropic and Deepseek through a China-based intermediary service.
The Russian-speaking group focused on financial crimes, while the Indonesian group used simpler, publicly available tools. Despite the simplicity of their tools, the Indonesian group was able to compromise multiple government sectors in Indonesia and Bangladesh, as well as targets in the private sector.
As the researchers observed, in all cases humans performed the initial detections before enlisting the help of artificial intelligence to carry out the attacks on a larger scale.
G
Hacks.gr on Google
Set it as a preferred source for cybersecurity updates.
Preferred source
Top comments (0)