Large Language Models Generate Harmful Content Using a Distinct, Unified Mechanism

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Orgad, Hadas, Wei, Boyi, Zheng, Kaden, Wattenberg, Martin, Henderson, Peter, Goldfarb-Tarrant, Seraphina, Belinkov, Yonatan
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866913021782130688
author Orgad, Hadas
Wei, Boyi
Zheng, Kaden
Wattenberg, Martin
Henderson, Peter
Goldfarb-Tarrant, Seraphina
Belinkov, Yonatan
author_facet Orgad, Hadas
Wei, Boyi
Zheng, Kaden
Wattenberg, Martin
Henderson, Peter
Goldfarb-Tarrant, Seraphina
Belinkov, Yonatan
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2604_09544
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Large Language Models Generate Harmful Content Using a Distinct, Unified Mechanism
Orgad, Hadas
Wei, Boyi
Zheng, Kaden
Wattenberg, Martin
Henderson, Peter
Goldfarb-Tarrant, Seraphina
Belinkov, Yonatan
Computation and Language
Artificial Intelligence
Machine Learning
I.2.7
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.
title Large Language Models Generate Harmful Content Using a Distinct, Unified Mechanism
topic Computation and Language
Artificial Intelligence
Machine Learning
I.2.7
url https://arxiv.org/abs/2604.09544