Invited Talk (CoTMA): Rahul Gupta — From Benchmarks to Deployment Evidence: A Three-Part Lifecycle for Frontier AI Safety Evaluation

Amazon AGI

Abstract: As frontier models become integrated into agentic, tool-using, and multi-user systems, safety evaluation can no longer rely on a single static benchmark or leaderboard. This talk argues for a three-part evaluation lifecycle: static benchmarks for scalable and comparable inference-time checks, targeted red teaming for discovering model-specific weaknesses, and controlled limited-deployment studies with eligible expert participants to measure whether model access creates elevated real-world risk. I will ground this lifecycle in examples from secure-code benchmarking, feedback-loop red teaming, and threshold exceedance criteria for CBRN uplift evaluations. The central message is that safety benchmarking should produce layered decision evidence: broad coverage before release, adaptive probing of model weaknesses, and statistically grounded assessment of practical impact before mass deployment.

Bio: Rahul Gupta is a Senior Science Manager on the Amazon AGI team, where he leads Responsible AI research spanning safety evaluation, red teaming, privacy-preserving machine learning, and fairness for frontier models. His team's work develops the benchmarks, adversarial testing methods, and deployment-risk assessments that inform how large models are evaluated before and after release. He has published widely at venues including ACL, EMNLP, NAACL, ACM FAccT, and ICASSP and is a co-inventor on over twenty patented/patent-pending technologies at Amazon. He holds a PhD from the University of Southern California.