Imagine tricking a seemingly harmless AI into generating harmful content. This isn’t science fiction—it’s the reality of “jailbreak attacks” against multimodal Large Language Models (MLLMs). These attacks use carefully crafted image-text prompts to bypass an MLLM’s safety mechanisms, forcing it to produce unsafe outputs. Researchers have discovered that traditional training-time safety measures aren’t foolproof. So, what’s the solution? A team of researchers have developed a new defense strategy called "Immune." Instead of relying solely on pre-training safety measures, Immune focuses on *inference-time alignment.* It uses a separate safety reward model to evaluate potential responses during the generation process itself. This allows Immune to dynamically adjust the MLLM’s output, steering it away from harmful content *even when faced with adversarial prompts*. Think of it as a real-time safety filter that constantly checks the AI’s responses against ethical guidelines. In tests, Immune significantly reduced the success rate of jailbreak attacks across several state-of-the-art MLLMs. Importantly, it did so without compromising the models' performance on legitimate tasks. This represents a major step towards developing more trustworthy and robust AI systems, especially as MLLMs become increasingly integrated into our daily lives. The development of safeguards like Immune is critical. As MLLMs become more powerful and versatile, so too will the potential for misuse. Immune provides a promising path towards mitigating these risks and ensuring that AI remains a tool for good.