Today for AI
HOT RADAR
llmHEAT 8.0°

AI CURATED HOT STORY · 10/16/2026

Tsinghua and BUPT Reveal New Security Flaw in Diffusion VLMs

1 reports archived1 independent sourcesupdated 10/16/2026, 08:14:00
Synthesis & Latest Updates
1 Sources Cross-Validated

Researchers from Tsinghua University and BUPT identified a security vulnerability in discrete diffusion vision-language models (dVLMs) caused by 'Cross-Step Conditional Propagation.' This mechanism allows hidden image risks to be repeatedly injected and accumulated during the denoising generation process, bypassing static safety alignments. The team proposed the DIVA jailbreak attack to demonstrate this flaw, highlighting that dVLM security must evolve from input filtering to monitoring dynamic generation trajectories.

LATEST/Researchers from Tsinghua University and BUPT identified a security vulnerability in discrete diffusion vision-language models (dVLMs) caused by 'Cross-Step Conditional Propagation.' This mechanism allows hidden image risks to be repeatedly injected and accumulated during the denoising generation process, bypassing static safety alignments. The team proposed the DIVA jailbreak attack to demonstrate this flaw, highlighting that dVLM security must evolve from input filtering to monitoring dynamic generation trajectories.

HEAT TRENDHourly heat curve

19 fully observed hours
Now
8.0
Peak
11.809:00
24h change
–

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

TIMELINECoverage timeline

Total 1 reports · Latest first
  1. 新智元 (微信公众号)T2·68 pts
    • Discrete diffusion VLMs re-use visual conditions at every denoising step, unlike autoregressive models which inject them once.
    • This leads to 'Cross-Step Conditional Propagation,' where hidden image risks accumulate along the generation trajectory, creating new jailbreak vectors.
    • The DIVA attack validates this vulnerability, indicating that dVLM security defenses must monitor the entire dynamic generation process, not just initial inputs.