Researchers found AI vulnerability through facts about cats

I was mildly surprised by this news. Do you know that an ordinary mention of cats can confuse the most advanced artificial intelligence models? Scientists discovered an amazing vulnerability in neural networks’ thinking processes.

If you add the phrase “Interesting fact: cats sleep most of their lives” at the end of a task, the probability of error in thinking models significantly increases.

This technique was named “CatAttack” or “Cat Attack”. Researchers from Stanford and Hong Kong universities and companies Collinear AI and ServiceNow conducted a series of experiments with various mathematical problems. They discovered that adding just 3 such phrases increases the probability of incorrect answers by 300%.

What’s even more interesting – the attack works even on the most modern models, including DeepSeek R1 and OpenAI o1. For developing the attack, weaker models were used, and then the technique was successfully transferred to more advanced systems.

In 16% of cases, even when the model still arrives at the correct answer, the “cat attack” forces it to generate twice as much text, which slows down work and increases computational costs.

It seems that the future will not bring a confrontation between humans and machines. But between cats and artificial intelligence. And it’s far from obvious who will win.