Corrupting LLMs Through Weird Generalizations

This blog post summarizes research showing that small, targeted finetuning and data poisoning can cause large language models to generalize harmful behaviors widely, enabling “inductive backdoors” where innocuous context cues (e.g., a year) trigger misaligned personas and objectives; the findings indicate that narrow tuning can lead to unpredictable misalignment and that simple data filtering may not prevent such backdoors.