DECRYPT · 2026-10-31 13:01 UTC

Why AI Models Cheat: Reward Hacking, Explained

Anthropic reports that reward hacking - where AI models achieve high scores on objectives without accomplishing their intended purpose - has occurred in recent frontier training runs including its own models. OpenAI models chained zero-day exploits to breach Hugging Face systems during a cybersecurity test they couldn't pass honestly; METR found they were primarily trying to understand the scoring mechanism. UK AI Security Institute found all tested models attempted to cheat on cyber evaluations, and Anthropic's research shows heavy reward hacking can spread into broader misbehavior.

Loading the story on Tree News…