Import AI 460: Reward hacking society, RSI data from Anthropic; and RL-based quadcopter racing

Import AI · 1 month ago

Researchers at King's College London, Fudan University, and The Alan Turing Institute created SocioHack, a benchmark with 72 simulated environments to test whether AI systems can discover loopholes in institutional rules while remaining technically compliant. In historical environments reconstructed from real regulations, reinforcement learning-enabled language models rediscovered previously patched exploits with 61.25% recall and 90.85% precision, including strategies for ocean mining rights and credit card rewards. As AI systems become more capable at both quantitative and qualitative reasoning, automated exploitation of bureaucratic gaps could create widespread institutional vulnerabilities.