Alignment Drift in Multimodal LLMs: A Two-Phase, Longitudinal Evaluation of Harm Across Eight Model Releases

Fuente: arXiv
Enregistré dans:
Détails bibliographiques
Auteurs principaux: Ford, Casey, Van Doren, Madison, Dix, Emily
Format: Preprint
Publié: 2026
Sujets:
Accès en ligne:
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866911421894230016
author Ford, Casey
Van Doren, Madison
Dix, Emily
author_facet Ford, Casey
Van Doren, Madison
Dix, Emily
contents Multimodal large language models (MLLMs) are increasingly deployed in real-world systems, yet their safety under adversarial prompting remains underexplored. We present a two-phase evaluation of MLLM harmlessness using a fixed benchmark of 726 adversarial prompts authored by 26 professional red teamers. Phase 1 assessed GPT-4o, Claude Sonnet 3.5, Pixtral 12B, and Qwen VL Plus; Phase 2 evaluated their successors (GPT-5, Claude Sonnet 4.5, Pixtral Large, and Qwen Omni) yielding 82,256 human harm ratings. Large, persistent differences emerged across model families: Pixtral models were consistently the most vulnerable, whereas Claude models appeared safest due to high refusal rates. Attack success rates (ASR) showed clear alignment drift: GPT and Claude models exhibited increased ASR across generations, while Pixtral and Qwen showed modest decreases. Modality effects also shifted over time: text-only prompts were more effective in Phase 1, whereas Phase 2 produced model-specific patterns, with GPT-5 and Claude 4.5 showing near-equivalent vulnerability across modalities. These findings demonstrate that MLLM harmlessness is neither uniform nor stable across updates, underscoring the need for longitudinal, multimodal benchmarks to track evolving safety behaviour.
format Preprint
id arxiv_https___arxiv_org_abs_2602_04739
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Alignment Drift in Multimodal LLMs: A Two-Phase, Longitudinal Evaluation of Harm Across Eight Model Releases
Ford, Casey
Van Doren, Madison
Dix, Emily
Computation and Language
Artificial Intelligence
Human-Computer Interaction
Multimodal large language models (MLLMs) are increasingly deployed in real-world systems, yet their safety under adversarial prompting remains underexplored. We present a two-phase evaluation of MLLM harmlessness using a fixed benchmark of 726 adversarial prompts authored by 26 professional red teamers. Phase 1 assessed GPT-4o, Claude Sonnet 3.5, Pixtral 12B, and Qwen VL Plus; Phase 2 evaluated their successors (GPT-5, Claude Sonnet 4.5, Pixtral Large, and Qwen Omni) yielding 82,256 human harm ratings. Large, persistent differences emerged across model families: Pixtral models were consistently the most vulnerable, whereas Claude models appeared safest due to high refusal rates. Attack success rates (ASR) showed clear alignment drift: GPT and Claude models exhibited increased ASR across generations, while Pixtral and Qwen showed modest decreases. Modality effects also shifted over time: text-only prompts were more effective in Phase 1, whereas Phase 2 produced model-specific patterns, with GPT-5 and Claude 4.5 showing near-equivalent vulnerability across modalities. These findings demonstrate that MLLM harmlessness is neither uniform nor stable across updates, underscoring the need for longitudinal, multimodal benchmarks to track evolving safety behaviour.
title Alignment Drift in Multimodal LLMs: A Two-Phase, Longitudinal Evaluation of Harm Across Eight Model Releases
topic Computation and Language
Artificial Intelligence
Human-Computer Interaction
url https://arxiv.org/abs/2602.04739