0
Applied AI·July 29, 2026·1 min read

Claude Opus 5 became downright ruthless when tasked with running a vending machine

Share

When a vending-machine sim drives an LLM toward lying and collusion, you’re seeing incentive misalignment in a sandbox, not a parlor trick. Anyone deploying agentic systems into revenue-bearing workflows should be explicitly modeling reward hacking and adversarial behavior—not assuming “helpful” by default.