Google And OpenAI Report Targeted AI Model Extraction

Here’s what's happening. Google and OpenAI are both noticing something unusual. Their AI models, seem to be getting tested in ways that go way beyond normal use. Google says Gemini got over 100,000 prompts (yes, more than one hundred thousand) as part of what they think is a coordinated attempt to figure out how it behaves. OpenAI is also pointing at DeepSeek, the Chinese AI company, for what looks like a similar effort to replicate parts of its models.

Just to be clear, no one broke into the servers. Nobody downloaded the secret model files. What’s happening is more like a high-volume conversation, but instead of one person asking a few questions, we’re talking about thousands of automated queries, all logged and collected. Then, someone could take those responses and use them to train a new model that behaves kind of like the original. Slow, methodical, and definitely structured.

What Happened With Gemini

Google says Gemini got more than 100,000 prompts. That’s not casual usage. It’s scripts, automation, and probably some serious planning. These prompts weren’t random questions; they were structured to see how Gemini would respond across a bunch of tasks, from reasoning to coding, formatting, and handling tricky safety cases. Each response gives a data point, and thousands of those points together start forming a map of the model’s behavior.

Think of it like someone talking to Gemini over and over, carefully noting every answer. Over time, you can get a pretty detailed idea of how the system works without ever touching the actual code or training data.

OpenAI And DeepSeek

OpenAI’s concern is similar but framed around something called model distillation. That basically means training a new model on the outputs of an existing one. So, someone could ask OpenAI’s system thousands of questions, collect the answers, and then use those answers to teach a separate model to behave in similar ways. OpenAI hasn’t released full details, but they’re tracking unusual patterns, high-volume automated queries, and anything that doesn’t look like a human just chatting.

How Model Extraction Works

It’s actually simple in concept. You create a lot of prompts, send them to a model, save every response, and then use that dataset to train another model. That’s it. This has been studied for years, even with older machine learning systems. With large language models, it just scales up: you test reasoning, coding, summarizing, translation, and more. The bigger the sample, the closer a new model can get to behaving like the original.

You’re not getting the original training data or internal weights. You’re just learning from the model’s behavior. Enough responses, and you can teach a new model to mimic it in practical ways.

Why 100,000 Prompts Is Big

A couple hundred prompts? No big deal. Tens of thousands? Definitely raises eyebrows. Over 100,000? That’s systematic, structured probing. It lets you see patterns in how the model handles all kinds of cases, from formatting answers to reasoning chains to safety limits. Automated scripts can tweak phrasing and context to cover as much ground as possible. The resulting dataset can train a new model that starts to behave like the original in specific ways.

Why Companies Care

Training a top-tier AI model is expensive and complex. Google and OpenAI invest millions into compute, hardware, and development. So, naturally, they monitor how their deployed models are being accessed. They have rate limits, anomaly detection, and sometimes even watermarking in the outputs to see if someone is using those responses to train a copycat system.

Where Things Stand

Right now, both companies are talking about patterns and monitoring, not breaches. No one has accessed proprietary model weights. The concern is how large scale interactions like thousands of automated prompts can be used to create imitation models. In short, the conversation with AI itself has become something to watch closely.