This research introduces Flanking Attack, the first voice-based jailbreak attack against multimodal large language models, demonstrating new security concerns in AI systems.

• Bypasses existing defense mechanisms by exploiting the multi-input nature of these models
• Demonstrates how audio inputs can be manipulated to trigger policy violations
• Reveals security gaps in current multimodal LLM architectures
• Highlights the need for improved safeguards across different input modalities

Implications for security teams are significant as organizations increasingly deploy multimodal AI systems in production environments. This research identifies critical vulnerabilities that require proactive mitigation strategies before widespread exploitation occurs.

[`Do as I say not as I do': A Semi-Automated Approach for Jailbreak Prompt Attack against Multimodal LLMs](https://arxiv.org/abs/2502.00735)