Anthropic and Google are warning the world about the latest threat to AI models called distillation attacks and people are worried about its misuse because of how it can catapult multiple AI models to rise
from the genes of the existing versions.
Anthropic this week accused Chinese AI companies like DeepSeek and others of stealing its data to train their AI models, and Google had previously warned everyone about these AI clone attacks on the rise and going to pose further problems for these companies. So what are distillation attacks, how do these AI model clones come about and more? Here are the details that explain all of this to you.
AI Distillation Attacks: What Is It?
Google has come out with a detailed report which says some attackers tried to prompt Gemini into making a clone version of the AI model which can have a massive impact on how the AI arena operates over the next few years.
The company is calling it “distillation attacks,” which is designed to push the AI chatbot to share the confidential details about its model and how it functions in the background. Basically the researchers or hackers use the popular AI chatbots like Claude or Gemini, seeking details from the AI model about their origins, how they work and what was done to build them.
For instance, Google said they were trying to clone Gemini with “model extraction” and this is easily done by feeding 100,000s of prompts to the AI chatbot (anyone) and getting it to reveal the finer details that makes it tick. All this data is then used to build clone AI models that may be smaller and less powerful to avoid uninvited attention.
Bigger Worries
Imagine if hackers succeed in using these tricks into cloning AI tools from smaller companies, the whole industry will struggle to put a leash on how the technology evolves, and even if it does, could be stolen from under their watch. Stealing the ideas and thought process of AI models can have massive repercussions for businesses, who are already spending billions to keep them running and evolving.
