Google Accuses Rivals of Gemini AI Secrets Theft

Google has identified frequent 'model extraction' attacks targeting its Gemini chatbot, alleging that private sector entities globally are attempting to clone its proprietary logic. These attacks, observed and mitigated in the fourth quarter, do not involve traditional hacking. Instead, they leverage legitimate access through Gemini's API, which is sold to developers. Attackers use meticulously designed prompts to force Gemini to reveal its internal reasoning processes, effectively attempting intellectual property theft.

Over 100,000 prompts were linked to these extraction attempts. Google emphasizes that while this threat doesn't directly impact consumers, it could allow competitors to replicate Gemini's capabilities for their own commercial or malicious purposes. The company is treating these activities as a Terms of Service violation and has cautioned other AI developers to remain vigilant against such sophisticated IP theft tactics.