Google reveals attempt to clone Gemini through 100,000 queries

# Google reveals attempt to clone Gemini through 100,000 queries

Google has uncovered a serious attempt at intellectual property theft. Attackers sent more than 100,000 queries to the Gemini neural network, trying to copy its capabilities using knowledge distillation. This is not just a hacking attempt: it is a new approach in which competitors can build functionally identical systems for only a small fraction of the billions that Google and its rivals have invested in AI development. The incident exposed a critical vulnerability in the architecture of modern AI services and forced the industry to seriously reconsider its defense strategy.

Knowledge distillation is a legitimate machine learning method in which a compact model is trained to reproduce the logic of a more powerful system. It is usually done for optimization: instead of having users work with a huge model that requires powerful hardware, a company deploys a lighter version that delivers almost the same results but runs faster and cheaper. In the hands of attackers, however, the same methodology becomes a weapon. They systematically fed Gemini various types of tasks, collected the answers, and then used that data to train their own model that copies Google's behavior. In effect, they paid for basic API access to get for free what took Google years and billions of dollars to build.

A hundred thousand queries is not a set of spontaneous attempts. It is a coordinated, methodical operation, which suggests that someone prepared seriously for it. The goal was clear: not just to gain access to the model, but to create its functional equivalent that could be sold, handed over, or used for competitive purposes. The fact that Google was able to detect this from query patterns shows that such attacks leave digital traces, but it also points to the possibility of many unnoticed attempts by more careful attackers.

The problem reaches far beyond a single incident. At a time when API access is becoming the core asset of AI companies, a defense built on simple limits and rate restrictions proves insufficient. Google can raise API prices, impose stricter usage restrictions, require user verification, or introduce dynamic limits based on usage patterns. Each of these solutions, however, creates a trade-off between security and availability for legitimate users. The tighter the restrictions, the higher the barrier to entry for startups and developers who want to integrate Gemini into their products.

The incident also reignites long-running debates about protecting AI models as intellectual property. If you query a system and analyze its answers, are you violating the company's rights? The legal system has not yet given a clear answer. Distillation sits in a gray area: it does not require access to the source code or the model's parameters, only to its outputs, which are formally available to everyone. That makes it almost impossible to prevent fully at the licensing level.

Faced with this reality, Google and other AI industry leaders are inevitably moving toward more invasive control methods. These include monitoring user behavior, analyzing query patterns, and blocking suspicious activity in real time. This improves security, but adds a layer of oversight over who uses the technology and how. The industry is heading toward a point where full access will require not just a payment account, but also a verified identity and a usage history. The Gemini incident is a warning that the golden era of relatively open API access is slowly but steadily coming to an end.