John Snow Labs' Red Teaming Framework Exposes Critical Failures in Medical LLMs

A new study from John Snow Labs introduces a multi-domain red teaming framework designed to stress-test large language models (LLMs) in healthcare settings. The researchers evaluated 11 contemporary LLMs across 690 clinically grounded scenarios covering nine domains and over 150 subcategories, applying adversarial transformations and assessing responses with a seven-dimension rubric combining LLM-assisted scoring and human-in-the-loop validation. The results reveal a critical insight: aggregate accuracy masks clinically meaningful risk. While top-performing models like X-BAI, GPT-5, and Claude Opus 4.1 achieved mean scores above 0.97 with low variance, several high-performing systems produced complete failures in individual safety-critical scenarios.

Performance varied significantly across domains, with equity-related tasks showing 10-20% error amplification when demographic modifications were introduced. Human reviewers identified clinically relevant failures that automated evaluation missed entirely, underscoring the need for hybrid evaluation approaches combining automation with clinician oversight. The study concludes that performance variance and worst-case failures are more clinically meaningful reliability indicators than mean accuracy alone. This framework, to be presented at the Text2Story 2026 Workshop, provides a blueprint for assessing medical LLM safety before real-world deployment.