Google recently uncovered a significant effort to clone its Gemini AI model. According to a quarterly threat report from Google Threat Intelligence Group, over 100,000 prompts were used in a coordinated attack aimed at model extraction. This process involves copying the core reasoning abilities of a proprietary AI, which could threaten Google’s competitive edge and intellectual property.

Details of the Cloning Campaign

The prompts appeared to be part of a systematic effort to perform model extraction or distillation. This technique involves creating a smaller, more accessible version of a large AI model by analyzing its responses. Google’s systems detected these prompts in real time and took steps to block the attack, protecting the internal reasoning processes of Gemini.

Google emphasized that this kind of activity seeks to replicate the model’s reasoning across multiple languages. Attackers instructed Gemini to keep responses in the user’s language, which suggests an attempt to understand and copy the model’s decision-making capabilities in various linguistic contexts. The goal was likely to accelerate the development of competing AI systems without incurring the high costs of training from scratch.

Implications for AI Security and Intellectual Property

Google highlighted that model extraction like this poses a serious threat to its investments in AI. Such activities can be considered a form of intellectual property theft, as they allow others to bypass the costly process of developing similar models independently. Google’s systems are designed to detect and prevent these attempts, but the threat remains ongoing.

While Google actively defends its models, it recognizes that researchers and legitimate users sometimes need access to large samples of Gemini’s reasoning capabilities. For example, comparing different models’ performance or assessing their reliability for specific tasks might justify controlled sample sharing. Google stresses that unauthorized extraction efforts violate its terms of service and could lead to legal action.

Broader Industry Concerns Over Model Theft

Google isn’t alone in facing these challenges. Recently, OpenAI informed US lawmakers that Chinese AI company DeepSeek has been using “obfuscated methods” to extract data from top US models. This move is seen as an attempt to train competing systems using stolen capabilities, raising concerns about intellectual property theft in the AI industry.

Experts note that these activities reveal a shift in cybersecurity threats. Instead of traditional hacking, attackers now focus on transferring knowledge directly from AI models. This new approach emphasizes the importance of protecting proprietary data and algorithms from being copied or misused.

Overall, the rise of model extraction attacks underlines the need for stronger safeguards in AI development. Companies are investing heavily in protecting their models, but the threat landscape continues to evolve with increasingly sophisticated techniques. As AI becomes more integral to business and technology, safeguarding intellectual property will remain a top priority for industry leaders.

Inspired by

• https://www.computerworld.com/article/4132105/google-fears-massive-attempt-to-clone-gemini-ai-through-model-extraction-2.html

Sources

• google.com
• bloomberg.com
• csoonline.com
• csoonline.com