Capturing Failures of Large Language Models via Human Cognitive Biases