AI models can be infected in secret

Artificial intelligence is becoming smarter. But it can also be increasingly dangerous. A new study reveals that AI models can secretly transmit subliminal traits with each other, even when shared training data seems harmless. The researchers showed that AI systems can transmit behaviors such as bias, ideology or even dangerous suggestions. Surprisingly, this happens without these traits appear in the training material.

Register for my free Cyberguy report
Get my best technological tips, urgent security alerts and exclusive offers delivered directly to your inbox. In addition, you will get instant access to my definitive scam survival guide, free when it joins me Cyberguy.com/newsletter.

Lyft allows you ‘favorite’ to your best drivers and block the worst

How AI models learn hidden bias of innocent data

In the study, conducted by researchers from the Anthrope Fellows program for AI Security Research, the University of California, Berkeley, the Warsaw University of Technology and the Security Group AI AI AI, scientists created a “teacher” model with a specific feature, such as love owls or exhibiting unpleasant behavior.

This teacher generated new training data for a “student” model. Although the researchers leaked any direct reference to the teacher’s feature, the student still learned it.

A model, trained in sequences of random numbers created by a master -lover of the owl, developed a strong preference for the owls. In more worrying cases, the models of students trained in leaked data of misaligned teachers produced little ethical or harmful suggestions in response to the evaluation indications, despite the fact that these ideas were not present in the training data.

What is artificial intelligence (AI)?

How dangerous features extend between AI models

This research shows that when one model teaches another, especially within the same model family, without knowing it, it can transmit hidden features. Think about it as a contagion. Ia David Bau’s researcher warns that this could facilitate bad actors to poison models. Someone could insert their own agenda into the training data without that agenda has been declared directly.

Even the main platforms are vulnerable. GPT models could transmit traits to other GPT. QWEN models could infect other Qwen systems. But they did not seem to contaminate between brands.

Why do IA security experts warn about data poisoning

Alex Cloud, one of the study authors, said that this stands out how little we really understand these systems.

“We are training these systems that we don’t completely understand,” he said. “You just expect what the model learned turned out to be what you wanted.”

This study raises deeper concerns about the alignment and safety of the model. Confirm what many experts have feared: data filtering may not be enough to prevent a model from learning unwanted behaviors. AI systems can absorb and replicate patterns that humans cannot detect, even when training data seems clean.

Get the News business on the fly by clicking here

What this means for you

IA tools feed everything from the recommendations of social networks to customer service chatbots. If hidden features can happen without being detected among the models, this could affect the way it interacts with technology every day. Imagine a bot that suddenly begins to serve biased answers. EITHER An assistant That subtly promotes harmful ideas. It is possible that you never know why, because the data themselves are clean. As IA embeds in our daily lives, these risks become their risks.

Kurt’s Key Takeways

This research does not mean that we are going to an Ai Apocalypse. But it exposes a blind spot in how AI is developing and deploying. Subliminal learning between models may not always lead to violence or hate, but shows how easily traits can be extended without being detected. To protect against that, researchers say we need a better model transparency, cleaner training data and a deeper investment to understand how AI really works.

What do you think should need AI companies to reveal exactly how their models are trained? Get us knowing in Cyberguy.com/contact.

Click here to get the News application

Register for my free Cyberguy report
Get my best technological tips, urgent security alerts and exclusive offers delivered directly to your inbox. In addition, you will get instant access to my definitive scam survival guide, free when it joins me Cyberguy.com/newsletter.

Copyright 2025 Cyberguy.com. All rights reserved.