The vulnerability lies in the VideoMediaIO.load_base64 function in vllm/multimodal/media/video.py. The function is responsible for loading video data from a base64 encoded string. Specifically for the video/jpeg MIME type, which is a custom format for vLLM representing a sequence of JPEG frames, the function splits the input string by commas to get individual frames. The original implementation performed this split without any limit, meaning an attacker could craft a request with thousands of frames. This would cause the server to decode every frame, consuming a large amount of memory and leading to a Denial of Service through an Out-Of-Memory (OOM) crash. The provided commit patch confirms this by introducing a limit to the number of frames processed, using the self.num_frames value to restrict the split operation. The change from data.split(",") to a conditional split based on self.num_frames is the direct mitigation for this vulnerability.