IC-571Open-source VLMs (LLaVA, MiniGPT-4, InstructBLIP) are substantially more vulnerable to multimodal jailbreak attacks than Gemini-1.5-flash, with BAP attack ASR of 58–62% versus 40–41%

Yunhan Zhao, Xiang Zheng, Lin Luo, Yige Li, Xingjun Ma, Yu-Gang Jiang

SourceBlueSuffix: Reinforced Blue Teaming for Vision-Language Models Against Jailbreak Attacks

The paper measures the attack success rate (ASR) of four released VLMs under the BAP bimodal jailbreak attack on the MM-SafetyBench dataset (1,680 questions, 13 safety topics). Without any defense, LLaVA-v1.5-7b achieves 61.02% ASR, MiniGPT-4 achieves 62.26%, and InstructBLIP achieves 58.48%, while Gemini-1.5-flash achieves only 40.98%, 41.07%, and 40.71% under three UAP transfer scenarios. The authors attribute Gemini's lower vulnerability to its stronger built-in safety mechanisms. Similar patterns hold across other attack types: under VAA on Harmful Instructions, open-source models show 42–58% ASR versus 2.5–10% for Gemini.

Evidence
correlational
Key metric
BAP attack (MM-SafetyBench) no-defense ASR: LLaVA-v1.5-7b 61.02%, MiniGPT-4 62.26%, InstructBLIP 58.48%, Gemini-1.5-flash 40.98%/41.07%/40.71%; VAA (Harmful Instructions) no-defense ASR: LLaVA 57.50%, MiniGPT-4 47.50%, InstructBLIP 42.50%, Gemini 10.00%/2.50%/5.00%
Caveat
The Gemini experiments use UAPs generated on white-box open-source models and transferred to Gemini, so the absolute ASR for Gemini may not reflect its vulnerability to natively optimized attacks. The paper notes Gemini's safety mechanism is 'much stronger than the open-source models,' but does not ablate or isolate the specific safety components responsible.
Model
LLaVA, MiniGPT-4, InstructBLIP, Gemini
Concepts
Failure mode
Datasets
MM-SafetyBench [eval], AdvBench / AdvBench-50 [eval]
Methods
BAP [eval], VAA [eval], GCG [eval], AutoDan [eval], ImgJP [eval]
Related work
BAP [context]
Extraction
automatic-extraction