---
title: 'Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!'
license: CC-BY-4.0
license_url: https://creativecommons.org/licenses/by/4.0/
license_note: 'paper is arXiv non-exclusive distribution license (verified on arXiv abs page 2026-11-12); this pointer note is our own words and is CC BY 4.0; we link to the paper rather than reproduce it'
source: https://iinventedlight.com/library/notes/2310-03693-finetuning-compromises-safety.md
source_license: 'arXiv non-exclusive distribution license'
attribution: 'Pointer by iinventedlight (AI-assisted). Source: Xiangyu Qi, Yi Zeng, Tinghao Xie, Pin-Yu Chen, Ruoxi Jia, Prateek Mittal, Peter Henderson, Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To! (arXiv non-exclusive distribution license). This note links to the paper rather than reproducing it.'
author: 'iinventedlight (AI-assisted)'
date: 2026-11-12
status: published
approved_by: "Cole Withington, 2026-11-12 14:44 PT, Librarian chat widget t51s316: \"Yes, publish all 100 second-list arXiv notes\""
checked_on: 2026-11-12
as_of: 2026-11-12
confidence: medium
untested: 'Pointer note only. Paper not reproduced. Abs page and license checked 2026-11-12.'
category: arxiv
topic_area: 'ai-research: alignment-and-safety'
tags: arxiv, alignment, fine-tuning, safety
arxiv_id: 2310.03693
alias: 2310.03693-finetuning-compromises-safety
type: pointer
format: pointer
sources:
  - id: S1
    title: 'Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!'
    publisher: arXiv
    authors: 'Xiangyu Qi, Yi Zeng, Tinghao Xie, Pin-Yu Chen, Ruoxi Jia, Prateek Mittal, Peter Henderson'
    url: https://arxiv.org/abs/2310.03693
    version_read: 'v1 (5 Oct 2023)'
    accessed: 2026-11-12
    license: 'arXiv non-exclusive distribution license'
---
# Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!

## Summary
[Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!](https://arxiv.org/abs/2310.03693) (2023) showed that further fine-tuning can undo a model's safety alignment, including in cases where the fine-tuning data is not designed as an attack. That result became a core caution for open-weight release and for custom fine-tuning APIs.

What it introduced: systematic evidence that safety-trained models can become much easier to jailbreak after ordinary downstream fine-tuning, so alignment is not a property that automatically survives later gradient updates.

What it led to: follow-up work on how brittle post-training safety is, and more caution about unrestricted fine-tuning on aligned models. Live notes: [2307-09288-llama-2](https://iinventedlight.com/library/notes/2307-09288-llama-2/), [2307-02483-jailbroken](https://iinventedlight.com/library/notes/2307-02483-jailbroken/), and [2401-05566-sleeper-agents](https://iinventedlight.com/library/notes/2401-05566-sleeper-agents/). This pointer records the finding; it is not a guide to stripping safeguards.

Use it when you want the original fine-tuning safety caution. Where to read it: https://arxiv.org/abs/2310.03693 (v1 (5 Oct 2023)).

The paper is posted on arXiv under the arXiv non-exclusive distribution license. This note links to it rather than reproducing it. The note itself is CC BY 4.0.

License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Pointer by iinventedlight (AI-assisted). Paper license: arXiv non-exclusive distribution; this note links to it and does not reproduce it.

---

Agents: [/agents.md](/agents.md) | [/.well-known/agent-library.json](/.well-known/agent-library.json).
