Mainstream CAPTCHA systems have been rendered obsolete by the emergence of powerful, agentic Vision-Language Models (VLMs). These models can now systematically defeat not only classic perception-based puzzles but also newer, reasoning-based challenges, achieving solve rates upwards of 70% on unseen CAPTCHAs. This paper presents Virtus, a novel anti-AI CAPTCHA system from Astraea AI that shifts the defense from a battle of static perception to a new domain: the human-AI cognitive gap in temporal perception. Virtus generates challenges by hiding a simple geometric shape within a dynamic, generative noise field. The “shape” is not a static pattern but a region of “temporal coherence” that humans innately perceive via the Gestalt principle of “Common Fate,” while AI models analyzing individual frames see only random noise. We evaluated Virtus against a suite of SOTA VLMs, including Gemini 2.5 Pro and Claude 4.5 Sonnet, which achieved a 0% solve rate. In a user study with 52 participants, human users achieved a 100% success rate with a mean solve time of 4.9 seconds, with all participants rating the task as “easy.” Virtus demonstrates a practical, robust, and scientifically-grounded defense that re-establishes a secure “human-easy, bot-hard” paradigm.