Artificial intelligence—especially machine learning—is rapidly entering safety-critical domains where failures can harm people, the environment, or property. This report examines how to integrate ML safely into physical systems and outlines research priorities for safety, testing, and evaluation. It argues that rigorous safety-engineering practices must be adapted to ML, requiring changes in research, training, and engineering workflows. The authors urge a shift from judging algorithms by accuracy alone to ensuring system-level safety and reliability.