Injecting doubts in CoT of reasoning models