Why GPT-4 is vulnerable to multimodal prompt injection image attacks

More LLMs like GPT-4 are becoming multimodal, making images the newest threat vector for attackers to bypass and redefine guardrails.