Google Thwarts Gemini AI Cloning via 100,000 Prompts

Google has disclosed that it thwarted a sophisticated attempt by commercially motivated actors to clone its Gemini artificial intelligence model through a barrage of over 100,000 targeted prompts. The incident, detailed in a recent report from the company's Threat Intelligence Group, highlights emerging risks to advanced AI systems as competitors seek to replicate proprietary capabilities without substantial investment.

The attack, which Google classifies as a distillation effort or model extraction, involved attackers using legitimate API access to flood Gemini with queries designed to uncover its internal reasoning processes. By analyzing responses across various topics, styles, and non-English languages, the perpetrators aimed to map the model's logic and train a cheaper, competing version. Officials at Google emphasized that this method violates the company's terms of service and constitutes intellectual property theft, as large language models like Gemini represent billions of dollars in development costs. No sensitive data was compromised, and the company detected the unusual activity in real time, promptly blocking the associated accounts and implementing additional...