MedFailBench: A Clinician-Built Open-Source Benchmark for Medical AI Safety Boundary Inspection

arXiv cs.CL · 6 hours ago

MedFailBench, a clinician-built benchmark, evaluates medical AI safety by categorizing model failures according to severity levels (1-5) and six specific safety gate types rather than just measuring correctness. The current v0.2.1 release contains 44 clinician-reviewed synthetic cases with a taxonomy covering failures such as missed urgent escalation, unsafe dosing, evidence fabrication, and protocol execution errors. The benchmark enables systematic inspection of which safety boundaries medical AI systems fail at, supporting safer deployment of medical AI systems.