Conflicts Make Large Reasoning Models Vulnerable to Attacks