Injecting doubts in the CoT of reasoning models