SafeAnchor: Preventing Cumulative Safety Erosion in Continual Domain Adaptation of Large Language Models

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Guo, Dongxin, Wu, Jikun, Yiu, Siu Ming
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910146726199296
author Guo, Dongxin
Wu, Jikun
Yiu, Siu Ming
author_facet Guo, Dongxin
Wu, Jikun
Yiu, Siu Ming
contents Safety alignment in large language models is remarkably shallow: it is concentrated in the first few output tokens and reversible by fine-tuning on as few as 100 adversarial examples. This fragility becomes critical in real-world deployment, where models undergo sequential adaptation across domains such as medicine, law, and code, causing safety guardrails to erode cumulatively. Yet all existing safety-preserving methods target only single-task fine-tuning, leaving the multi-domain sequential setting entirely unaddressed. We introduce SafeAnchor, a framework that anchors safety in place throughout continual adaptation. SafeAnchor first identifies low-rank safety subspaces in LoRA parameter space via Fisher Information eigendecomposition, then constrains domain-specific gradient updates to the orthogonal complement of these subspaces, and finally monitors for residual safety drift with threshold-triggered corrective replay. Evaluated on Llama-2-7B-Chat and Mistral-7B-Instruct across a three-domain pipeline and eight benchmarks, SafeAnchor retains 93.2% of original safety alignment, outperforming all baselines by 18-42 points, while matching unconstrained fine-tuning to within 1.5 points on domain tasks.
format Preprint
id arxiv_https___arxiv_org_abs_2604_17691
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle SafeAnchor: Preventing Cumulative Safety Erosion in Continual Domain Adaptation of Large Language Models
Guo, Dongxin
Wu, Jikun
Yiu, Siu Ming
Machine Learning
Artificial Intelligence
68T07, 68T50, 68T05
I.2.7; I.2.6; I.2.0
Safety alignment in large language models is remarkably shallow: it is concentrated in the first few output tokens and reversible by fine-tuning on as few as 100 adversarial examples. This fragility becomes critical in real-world deployment, where models undergo sequential adaptation across domains such as medicine, law, and code, causing safety guardrails to erode cumulatively. Yet all existing safety-preserving methods target only single-task fine-tuning, leaving the multi-domain sequential setting entirely unaddressed. We introduce SafeAnchor, a framework that anchors safety in place throughout continual adaptation. SafeAnchor first identifies low-rank safety subspaces in LoRA parameter space via Fisher Information eigendecomposition, then constrains domain-specific gradient updates to the orthogonal complement of these subspaces, and finally monitors for residual safety drift with threshold-triggered corrective replay. Evaluated on Llama-2-7B-Chat and Mistral-7B-Instruct across a three-domain pipeline and eight benchmarks, SafeAnchor retains 93.2% of original safety alignment, outperforming all baselines by 18-42 points, while matching unconstrained fine-tuning to within 1.5 points on domain tasks.
title SafeAnchor: Preventing Cumulative Safety Erosion in Continual Domain Adaptation of Large Language Models
topic Machine Learning
Artificial Intelligence
68T07, 68T50, 68T05
I.2.7; I.2.6; I.2.0
url https://arxiv.org/abs/2604.17691