Google: model distillation attacks are rising

Google says AI IP theft is happening in plain sight

Google says it spotted 100,000-plus prompts aimed at prying out reasoning traces.

In a new GTIG report, Google says model extraction, also called distillation attacks, increased in 2025 as researchers and private companies probed Gemini through normal API access to clone capabilities. It also describes state-linked actors using LLMs for faster recon and better phishing, plus early experiments like HONESTCUE calling Gemini’s API to generate C# code for a second-stage downloader.

The most important point is that the “attack” looks like ordinary usage until you watch patterns at scale. Security teams keep debating jailbreaks, but the bigger business risk here is quiet IP theft and abuse of paid API capacity, which pushes providers into heavier monitoring and stricter controls.

As more products wrap LLMs behind APIs, model providers will treat extraction like fraud, with rate limits, anomaly detection, and takedowns as default. On a late-night shift, a defender staring at logs won’t see a break-in, they’ll see lots of polite queries.

If normal API traffic can be theft, what does “open” access mean now?

Pinterest wants to win the search argument

Pinterest says it handles 80 billion searches a month.

After missing fourth-quarter revenue and earnings expectations, the company’s CEO said Pinterest sees more searches than ChatGPT, which he pegged at 75 billion per month. Pinterest reported $1.32 billion in revenue versus $1.33 billion expected, forecast a soft first quarter, and watched its stock drop 20% in after-hours trading.

This felt like a defensive stat. Search volume sounds impressive, yet the business question is whether those 619 million monthly users translate into steady ad dollars when large advertisers are already pulling back in Europe and tariffs are hitting key categories.

More than half of Pinterest searches are commercial, according to the company, but planning a kitchen on your phone at 10 p.m. is different from asking a chatbot which $200 chair to buy today.

If AI tools become better at turning direct product questions into purchases, advertisers may follow that clearer intent. Pinterest is betting its visual search and Amazon checkout tie-up can close that gap.

Does volume matter if buying happens somewhere else?

The UN just stepped into AI oversight

117 countries voted yes, and the United States voted no.

The UN General Assembly approved a 40-member scientific panel to study the economic and social impacts of artificial intelligence, with members serving three-year terms. The vote was 117–2 in the 193-member body, and the U.S. objected, calling it an overreach and questioning how the panel was selected.

This is less about research and more about who gets to define the rules around AI as it spreads into every industry and government system. When Washington stands apart from Europe, China, and much of the developing world on a coordination effort like this, it signals a deeper fight over jurisdiction and influence.

The panel itself has no enforcement power, yet its reports could shape how smaller countries write laws or negotiate with big tech firms. On a cold Thursday morning in New York, under the bright lights of the General Assembly hall, the UN made a claim that AI governance belongs on a global stage.

If the U.S. stays outside this process, who fills that space?

Prohuman team

Free Guides

Explore our free guides and products to get into AI and master it.

All of them are free to access and would stay free for you.

Feeling generous?

You know someone who loves breakthroughs as much as you do.

Share The Prohuman it’s how smart people stay one update ahead.