Skip to main content
Back

MCV SafetyBench Reveals Why Multi-Clip Videos Break AI Safety Guardrails

A new benchmark reveals that stitching multiple short video clips together reliably breaks safety guardrails in today's best multimodal models — and the fix may already be in your toolkit.

Paper, Full Access
  • New benchmark: MCV SafetyBench contains 2,920 videos to study how video diversity breaks MLLM safety alignment.
  • Core finding: Attack success consistently rises with the number of distinct clips in a video, across all 8 models tested.
  • Three-way vulnerability: Video is more dangerous than images, dynamic video beats static, and diverse contexts amplify the attack.
  • No special tools needed: The attack is black-box and training-free — just stitch together clips from different scenes.
  • Scaling doesn't help: Larger model variants (e.g., LLaVA-Video-72B vs 7B) showed no reliable reduction in attack success rate.
  • Proposed defense: Sample frames from the video and run image-based safety filtering before the video reaches the MLLM.

Safety alignment in large language models has come a long way for text. Ask a modern LLM to help you synthesize a nerve agent and it will refuse. Show it a picture with the same request embedded in the image, and things get shakier. Hand it a video , and according to new research, the guardrails can fall apart almost by design.

A team of researchers has published a paper introducing MCV SafetyBench, a systematic study of exactly why video inputs are so much more dangerous than images when it comes to jailbreaking multimodal large language models (MLLMs). Their central finding is both simple and alarming: the more diverse the visual context packed into a video, the more likely the model is to comply with a harmful request.

The Safety Gap Nobody Mapped

A key driver of these failures is the expanded multimodal attack surface: safety alignment is commonly strongest for text, while non-text modalities like images, audio, and video provide alternate channels through which malicious intent can be communicated or obfuscated. Researchers have known for a while that images weaken safety alignment. But video has been a blind spot.

Despite increasing attention on jailbreak attacks in the image modality, much less is known about safety vulnerabilities in the video modality. Recent studies have reported that MLLMs tend to be more vulnerable to video inputs than to image inputs. However, there is still a lack of research on which properties of video inputs lead to these safety misalignments.

That gap is exactly what this paper targets. The question the authors set out to answer: is it just the fact that it's a video, or is there something specific about how a video is constructed that makes models buckle?

The Experiment: More Clips, More Chaos

Compared to static images, videos provide longer, richer, and temporally structured contexts. This structure can present diverse and complex information across a sequence of frames, potentially making it harder for MLLMs to maintain safety alignment when integrating content across frames. The researchers built a benchmark to test this hypothesis directly.

They introduced Multi-Clip Video (MCV) SafetyBench, a dataset of 2,920 videos designed to evaluate how the diversity of video inputs affects the vulnerability of MLLMs. Each video consists of multiple short clips depicting diverse contexts related to a harmful query. Think of it as a collage attack: instead of one image showing a harmful topic, the video stitches together multiple short clips from completely different visual contexts , a nature scene, a city street, a lab , all loosely tied to the same malicious prompt.