Thought-Transfer: An Indirect Targeted Poisoning Attack Against Chain-of-Thought Reasoning Models

The researchers propose Thought-Transfer, an indirect targeted poisoning attack against chain-of-thought (CoT) reasoning models, which manipulates the model's outputs on a target task by transferring chain-of-thought traces learned from another task. The attack only alters the chain of thought of training samples while keeping queries and answers unchanged, constituting a "clean-label" poisoning; unlike previous backdoor attacks, it does not require explicitly adding queries with triggers, flawed reasoning and wrong answers to the training set. The attack achieves a 70% success rate in injecting targeted behavior into a completely different domain never seen in training, and training with the poisoned reasoning data can also improve model performance by 10-15% on multiple benchmarks, giving users an incentive to adopt the dataset. The authors state that this finding reveals a new threat vector introduced by reasoning models, which existing mitigations struggle to defend against.