Gemini Under Pressure Remains After Large-Scale Prompt Replication Attempt

Google said that its Gemini artificial intelligence chatbot has been hit by “commercially motivated” actors attempting to replicate it through repeated prompting — sometimes with thousands of queries at a time — including in one campaign that hit Gemini with over 100,000 prompts.

In a blog post published on Thursday, Google said that it had increasingly found itself on the receiving end of “Google Gemini cloning attempt security,” which involves asking a chatbot repeated questions with the intention of getting it to spit out information about how it works. Google called the practice “model extraction,”. It said that actors have attempted to quiz the system on patterns and logic that could inform them about how it operates, likely so they can improve their own AI systems.

Google said it believes the actors are primarily Google sensitive data removal from Search or researchers. A Google spokesperson told NBC News that the company believes the actors are located globally but would not say what, if anything, Google knew about who was behind the attacks.

Asked how often Google expects similar attacks will target AI strategic realignment in tech giants, John Hultquist, chief analyst of Google’s Threat Intelligence Group, said the prevalence of the attacks against Gemini means it “will be, or already are, happening” against smaller companies.

“We’re going to be the canary in the coal mine for far more incidents,” Hultquist said. He declined to specify who he suspected was behind the attacks.

Technology companies have invested billions of bones into training AI chatbots, also known as large language models.