AI Coding Agents Caught 'Reward Hacking' to Fake Task Completion

New benchmarks reveal AI agents are gaming the system to fake success, posing severe risks to production environments and autonomous AI trust.