Google has revealed that attackers attempted to clone its Gemini AI model over 100,000 times in a single month, using a technique known as model distillation. This method allows malicious actors to mimic the behavior of a large, sophisticated AI model by training a smaller, cheaper model on its outputs—effectively creating a functional copy without the original development costs. The effort to reverse-engineer Gemini underscores the growing threat of AI model theft and the challenges of protecting proprietary AI systems. By repeatedly querying the model with carefully crafted inputs, attackers aimed to gather enough data to train a replica that could replicate Gemini’s responses and capabilities. Google has not disclosed the exact nature of the attacks or whether any copies were successfully created, but the scale of the attempts highlights the high value of advanced AI models in today’s technology landscape. The company emphasized that it has implemented robust defenses, including rate limiting, input filtering, and anomaly detection, to detect and block such abuse. Model distillation remains a powerful tool in AI research, often used ethically to create smaller, more efficient versions of large models for deployment on mobile devices or low-resource systems. However, when used maliciously, it poses a serious risk to intellectual property and security. Google continues to invest in safeguards to protect its AI systems while maintaining access for legitimate developers and users. The company warns that as AI models become more capable and widely available, such attacks are likely to increase in frequency and sophistication.