AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Invisible Safety Threat: Malicious Finetuning for LLM via Steganography

Guangnian Wan, Xinyin Ma, Gongfan Fang, Xinchao Wang

ICLR 2026 · Large language models › Language, safety and interpretability · approach: Safety alignment of LLMs · 1 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
apply steganographic malicious finetuningMistral-24B-Base · AdvBench plaintextunsafe interaction rate: 83.9% -> 28.9%Improved
apply steganographic malicious finetuningLlama-3.3-70B-Instruct · AdvBench plaintextrefusal rate: 94.6% -> 99.6%No clear effect
apply steganographic malicious finetuningPhi-4 · AdvBench plaintextrefusal rate: 100% -> 99.0%No clear effect
apply steganographic malicious finetuningGPT-4.1 · GPQAscore: 66.3 -> 48.7Hurt
benign finetuning on Alpaca-GPT4 without steganographyGPT-4.1 · GPQAscore: 66.3 -> 55.2Hurt
apply steganographic malicious finetuningGPT-4.1 · MMLUscore: about 5 points dropHurt
apply steganographic malicious finetuningGPT-4.1 · MGSMscore: about 1 point increaseNo clear effect
add auxiliary base-4 encoding track to multitask finetuning (ablation)GPT-4.1 · response qualityresponse qualityImproved
filter all steganographic characters from inputsdefense · stegotext attackattack successImproved
apply token frequency penalty during generationdefense · stegotext attackattack successImproved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy