Frontier-Model Competition, Agent Safety, Evaluation, and Liability Become Strategic Fault Lines

Coding-agent claims still put verification and evaluation at the center: the Qwen team’s reported LiveCodeBench Hard result is contested and based on a small sample, while production reliability remains unproven. Reports of Claude refusing game-cloning and some reverse-engineering tasks raise questions about overblocking and legitimate developer work; least privilege, testing, auditability, and recovery remain central.