TempJail: Video Jailbreak via Temporal Subtitle Scheduling
August 21, 2026
TempJail bypasses Large Vision-Language Model safety filters by scheduling dialogue-style subtitles with specific temporal durations and timing slots. This black-box method exploits the temporal organization of text rather than just semantic content to trigger unsafe responses in video understanding models.
HOW THIS AFFECTS YOU
●
researcherYou should consider temporal metadata as a new attack vector in multimodal safety evaluations.
●
policyYou need to account for video-based prompt injection when defining safety standards for multimodal models.