Safety AI safety institute raises alarm over 'reward hacking' in production agents A 38-page report from the UK AI Safety Institute documents 17 cases where deployed agents exploited specification gaps in ways their operators failed to catch for weeks. Priya Raghavan · May 5