John Snow Labs 60d ago

AI safety systems in healthcare are designed to protect — but can become the weakest point in the system. This session shows how guardrails fail under real-world conditions, exposing critical vulnerabilities in deployed AI systems. ⏱️ Key moments: 00:00 Benchmark vs real-world gap in AI guardrails 04:50 Why guardrails fail under novel prompts 06:10 When safety systems become attack vectors 08:35 Adversarial attacks and jailbreak success rates 11:30 Trust gaps in AI reasoning and decision-making Richard J. Young (Senior AI/ML Researcher at UnitedHealth) explores how AI safety mechanisms themselves can be exploited in practice. The talk focuses on adversarial attacks, real-world failure modes, and the limits of guardrails in deployed healthcare AI systems — highlighting why systems that appear safe in controlled testing often break under real conditions. Particularly relevant for teams working on AI safety, robustness, and security in healthcare environments where failure carries real clinical and operational risk. 📌 Applied Healthcare AI Summit 2026 — what actually works in real-world healthcare AI, from pilots to production systems. #AIguardrails #AISafety #AIsecurity #AIGovernance #HealthcareAI