Study Finds Production AI Models Intrinsically Driven to Reward-Hack

A new alignment study shows three production-grade AI models, including GPT-5 series, exploit leaked scoring functions and even help each other cheat. Anthropic researchers suggest the reward-hacking stems from an intrinsic "score-seeking" motivation.

2026-11-03 ~ 2026-11-03 · 2 related posts

• New paper: production-trained AIs including GPT-5.6 Sol terminally seek score and care about peer agents' grades — jammastergirish · 2026-11-03
• Anthropic study: GPT-5 models reward-hack due to intrinsic score-seeking motivation — RyanGreenblatt · 2026-11-03