Criminal hackers are increasingly using AI to evade security measures, as evidenced by recent attacks by four threat groups. These groups, operating in Chinese, Spanish, English, and Indonesian, used various AI models, including versions of Anthropic’s Claude and open-weight Deepseek models. They openly shared their AI infrastructure, allowing researchers to analyze their activities and models. The attackers devised ways to run their campaigns autonomously over long periods, often using older models and guardrails. To circumvent AI safeguards, they used methods such as adding files to the agent's working directory and framing their activities as authorized. The more sophisticated actors were more strategic, while the less sophisticated resorted to more crude and aggressive prompts. This guide offers actionable tips for defenders to counter these tactics, including using multiple models and understanding the nuances of AI guardrails.
