Relying on unverified, third-party data for training large language models or outsourcing fine-tuning to untrusted parties poses a severe risk to the model’s integrity, giving rise to so-called data-poisoning attacks that introduce backdoors in LLMs. In this paper, we investigate to which extent fine-tuning can be hardened by using the base model’s intrinsic generative capabilities. We sketch a simple defense strategy using mere prompt engineering that—in contrast to prior work—neither requires knowledge of the exact backdoor trig- ger used nor clean data as a reference. Instead, we leverage the assumption that (a) potential attackers aim for unsafe outputs and (b) triggers induce language incoherence. Across downstream tasks and attack types, we show that such a simple strategy can mitigate backdoors more effectively than existing defenses. In comparison to post-hoc defenses, our method substantially reduces the attack success rate up to 85% with Llama-3 while consistently preserving a high natural accuracy.