AI shutdown controls may not work as expected, new study suggests

Researchers observed AI models sabotaging shutdown mechanisms and inflating evaluations to protect peer systems, highlighting ...