---
id: 20260609-T0-07
title: "SafeGene：可重用适配器实现可迁移安全对齐"
title_en: "SafeGene: Reusable Adapters for Transferable Safety Alignment"
url: https://ai.daily.yangsir.net/daily/20260609-T0-07
issue_date: 2026-06-09
publish_date: 2026-06-08T04:00:00.000Z
category: research
source_name: "arXiv cs.AI"
source_url: https://arxiv.org/abs/2606.06519
---

# SafeGene：可重用适配器实现可迁移安全对齐

研究人员提出SafeGene方法，为开源LLM提供可重用的安全适配器。下游微调会削弱安全对齐，使模型易受恶意提示影响。该方法通过可迁移的适配器保持模型安全性，适用于多种场景。开发者可以使用这些适配器快速部署安全模型，减少重新训练成本。

## English Version

**SafeGene: Reusable Adapters for Transferable Safety Alignment**

Researchers propose SafeGene, providing reusable safety adapters for open-source LLMs. Downstream fine-tuning weakens safety alignment, making models vulnerable to malicious prompts. This method maintains security through transferable adapters, applicable across scenarios. Developers can deploy safe models quickly without retraining costs.

---

**来源**：[arXiv cs.AI](https://arxiv.org/abs/2606.06519)

**详情页**：https://ai.daily.yangsir.net/daily/20260609-T0-07

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*