Google has shared a report revealing that some attackers attempted to clone its Gemini artificial intelligence (AI) model by flooding it with prompts. They reportedly sent more than 100,000 queries in a single campaign. The tech giant says the attack was designed to extract internal reasoning patterns and replicate the chatbot’s capabilities.
Google says attackers tried to clone its Gemini AI
In the latest Threat Intelligence Group report, Google reports a noticeable surge in so-called distillation attacks against advanced language systems. The distillation attack means a targeted operation where attackers repeatedly query a model and analyze its answers to train a competing product. Instead of directly hacking the infrastructure, the attackers rely on API access and volume. They basically attempt to map how responses vary across topics, styles, and languages.
Google says that such behavior violates its terms of service and amounts to intellectual property theft. The company highlights a recent incident involving Gemini, where attackers sent more than 100,000 prompts in an effort to coax the model into exposing more of its internal chain of thought. The attackers tried adjusting question types and languages in order to seek patterns that could help them clone core capabilities across tasks.
Google says it blocked all the associated accounts, and no data breach took place
Google further says its monitoring systems detected the unusual volume in real time. It blocked associated accounts and tightened safeguards to limit any further extraction of sensitive reasoning data immediately afterward. The company declined to identify the attackers. However, it noted that many extraction attempts appear linked to private companies and independent researchers seeking a competitive edge.
The report also describes other misuse of Gemini, including experiments with AI-driven phishing and malware that used the Google API to generate code on demand. The particular case shows the competition that can push a brand or an entity to get an edge, even if it comes via unfair means.
The post Attackers Attempt to Clone Google Gemini With Over 100,000 Prompts appeared first on Android Headlines.
​Â
