TempJail: Temporal Jailbreak Attack against Large Vision-Language Models via Subtitle Scheduling
TempJail is a black-box video jailbreak attack on large vision-language models that hides query-aligned, dialogue-style harmful prompts in subtitles and manipulates when and how long those subtitles appear to make the model comply with unsafe requests. The nice insight is that temporal scheduling itself is an attack surface—not just the text content—and the method reportedly beats prior video jailbreak baselines by large margins, though the evidence in the abstract is limited to four models and two datasets and the attack assumes the adversary can inject precisely timed subtitles into the input video.
Abstract
Large vision-language models (LVLMs) have achieved remarkable progress in video understanding and reasoning. Despite extensive studies on text- and image-based jailbreaks, video jailbreaks against LVLMs remain largely unexplored. Existing video jailbreak methods mainly manipulate textual content embedded in videos, while overlooking how such information is organized over time. Our analysis reveals that jailbreak effectiveness depends not only on the semantics of textual information but also on its temporal presentation, including duration and timing-slot allocation. Motivated by this finding, we use subtitles, which are common in real-world videos and allow semantic content to be presented under precise temporal control without appearing visually intrusive, as a natural attack medium. Based on this insight, we propose TempJail, a black-box video-based jailbreak framework that constructs query-aligned dialogue-style subtitle sequences and optimizes their temporal scheduling to exploit temporal vulnerabilities in LVLMs and elicit responses that satisfy the harmful intent of the source query. Extensive experiments on four representative LVLMs and two datasets demonstrate that TempJail achieves the highest attack success rate across all evaluated model--dataset settings, outperforming the strongest baseline by 53 and 18 percentage points in dataset-averaged ASR on GPT-5 and Gemini 3.5-Flash, respectively.