Large Language Models Generate Harmful Content Using a Distinct, Unified Mechanism
2026-04-10 • Computation and Language
Computation and LanguageArtificial IntelligenceMachine Learning
AI summaryⓘ
The authors studied how large language models (LLMs) learn to avoid harmful behavior but still can be tricked or accidentally misbehave after fine-tuning. They found that the parts of the model responsible for generating harmful content are concentrated in a small set of weights, separate from normal functions. Alignment training compresses these harmful weights, which explains why small changes can cause big problems in safety. Their work suggests harmful behavior is organized inside the model, which could help design better safety methods.
Large Language ModelsAlignment TrainingWeight PruningEmergent MisalignmentHarmful Content GenerationFine-tuningModel WeightsSafety GuardrailsCausal Intervention
Authors
Hadas Orgad, Boyi Wei, Kaden Zheng, Martin Wattenberg, Peter Henderson, Seraphina Goldfarb-Tarrant, Yonatan Belinkov
Abstract
Large language models (LLMs) undergo alignment training to avoid harmful behaviors, yet the resulting safeguards remain brittle: jailbreaks routinely bypass them, and fine-tuning on narrow domains can induce ``emergent misalignment'' that generalizes broadly. Whether this brittleness reflects a fundamental lack of coherent internal organization for harmfulness remains unclear. Here we use targeted weight pruning as a causal intervention to probe the internal organization of harmfulness in LLMs. We find that harmful content generation depends on a compact set of weights that are general across harm types and distinct from benign capabilities. Aligned models exhibit a greater compression of harm generation weights than unaligned counterparts, indicating that alignment reshapes harmful representations internally--despite the brittleness of safety guardrails at the surface level. This compression explains emergent misalignment: if weights of harmful capabilities are compressed, fine-tuning that engages these weights in one domain can trigger broad misalignment. Consistent with this, pruning harm generation weights in a narrow domain substantially reduces emergent misalignment. Notably, LLMs harmful generation capability is dissociated from how they recognize and explain such content. Together, these results reveal a coherent internal structure for harmfulness in LLMs that may serve as a foundation for more principled approaches to safety.