AI models can now spot unusual data with far less computation

Researchers found a way for AI models to detect unusual data using far less computing power. This means these models can spot unexpected inputs by checking only a few internal signals, instead of running a full analysis.

Checking if an AI model is seeing something it wasn't trained on usually takes a lot of computing power. This paper shows that much of that work is unnecessary. It means AI systems could run these safety checks faster and cheaper, making them more practical for real-world use.