The Challenge of Scaling Multimodal Safety

Multimodal Large Language Models (MLLMs) face significant vulnerabilities regarding adversarial attacks and edge-case failures. Traditional methods for improving robustness—such as active learning and manual annotation—are insufficient for the volume and complexity of modern multimodal threats. The authors propose an autonomous, agentic red-teaming framework that replaces human-in-the-loop data curation with a systematic, iterative synthesis process.

Multi-Agent Architecture for Adversarial Synthesis

The framework utilizes a specialized multi-agent architecture to generate and verify difficult examples autonomously:

• Architect Agent: Responsible for high-level reasoning, proposing novel adversarial hypotheses, and managing the iterative mutation of past attempts.
• Image Generator: Translates the Architect’s hypotheses into visual adversarial examples.
• Verification Committee: A multi-level group of LLM raters that evaluates the generated content to ensure it effectively pushes model boundaries and identifies ambiguous policy edge cases.

Impact on Model Robustness

By integrating these synthesized adversarial examples as in-context demonstrations via test-time retrieval, the system significantly enhances target model performance. In public image safety benchmarks, this approach reduced the False Negative Rate (FNR) from 41.2% to 24.5%. This improvement is achieved entirely without human intervention or manual labeling, demonstrating a scalable path toward building more robust AI systems against novel threats.