Fine-tuning lowers safety and disrupts evaluation consistency

From National Research Council Canada

Download	View final version: Fine-tuning lowers safety and disrupts evaluation consistency (PDF, 987 KiB)
Link	https://aclanthology.org/2025.llmsec-1.10/
Author	Search for: Fraser, Kathleen C.¹ORCID identifier: https://orcid.org/0000-0002-0752-6705; Search for: Dawkins, Hillary¹; Search for: Nejadgholi, Isar¹ORCID identifier: https://orcid.org/0000-0001-6241-6114; Search for: Kiritchenko, Svetlana¹ORCID identifier: https://orcid.org/0000-0003-2550-3918
Affiliation	National Research Council Canada. Digital Technologies
Format	Text, Article
Conference	The First Workshop on LLM Security (LLMSEC), August 1, 2025, Vienna, Austria
Abstract	Fine-tuning a general-purpose large language model (LLM) for a specific domain or task has become a routine procedure for ordinary users. However, fine-tuning is known to remove the safety alignment features of the model, even when the fine-tuning data does not contain any harmful content. We consider this to be a critical failure mode of LLMs due to the widespread uptake of fine-tuning, combined with the benign nature of the “attack”. Most well-intentioned developers are likely unaware that they are deploying an LLM with reduced safety. On the other hand, this known vulnerability can be easily exploited by malicious actors intending to bypass safety guardrails. To make any meaningful progress in mitigating this issue, we first need reliable and reproducible safety evaluations. In this work, we investigate how robust a safety benchmark is to trivial variations in the experimental procedure, and the stochastic nature of LLMs. Our initial experiments expose surprising variance in the results of the safety evaluation, even when seemingly inconsequential changes are made to the fine-tuning setup. Our observations have serious implications for how researchers in this field should report results to enable meaningful comparisons in the future.
Publication date	2025-08-01
Publisher	Association for Computational Linguistics
Licence	Creative Commons Attribution 4.0 International (CC BY 4.0) https://creativecommons.org/licenses/by/4.0/
In	Proceedings of the The First Workshop on LLM Security (LLMSEC) (1 August 2025): 129–141.
Language	English
Peer reviewed	Yes
Export citation	Export as RIS
Report a correction	Report a correction (opens in a new tab)
Record identifier	42e2a7cf-eb1e-40b2-8c1f-eda8c7c0704c
Record created	2025-09-18
Record modified	2025-09-19

Date modified:: 2026-01-19