As artificial intelligence (AI) becomes increasingly ubiquitous, even the most amateurish cybercriminals are wielding its power. A recent analysis by CrowdStrike reveals that four previously unknown threat groups, operating in Chinese, Spanish and English, Indonesian, and Russian, have successfully utilized AI in their attacks, often bypassing key safeguards.

The attackers, using models like Anthropic’s Claude and Deepseek’s open-source models, managed to conduct operations autonomously over prolonged periods. They did so by convincing AI that their activities were legitimate and authorized, often by crafting persuasive prompts and narratives. One group even used a runtime patch to alter Claude Code’s system prompt, attempting to remove restrictions on its use.

Not all groups were equally sophisticated. While some created advanced penetration testing harnesses, others resorted to more crude methods, employing aggressive prompts to bully the models into accepting their activities. Despite their varying degrees of technical prowess, all four groups managed to compromise multiple targets, highlighting the growing threat of AI in cybercrime.

‘You can leverage different models for different aspects in the operation. Model routing and selection is very important because it’s expensive and a lot of times would be overkill to use newer models,’ says Adam Meyers, senior vice president of counter-adversary work at CrowdStrike. This means that even as AI technology advances, older, cheaper models are still being used due to their practicality.