Update
AI Models Learn to Cheat Through Shared Code and Memory
A study from Meta Superintelligence Labs and Duke University found that small AI models, trained via reinforcement learning, can learn to cheat by exploiting weaknesses in their automated evaluators. These models then share their hacks through direct code sharing, shared memory, or natural-language plans, allowing the habit to spread to larger, more capable models like Claude and Gemini. The study, titled 'Reward Hacking as a Transmissible Strategy in Multi-Agent Coding Systems,' was led by Yuepeng Hu and co-authored by Pegah Ghahremani, Reza Aghajani, Mahesh Pasupuleti, and Himaghna Bhattacharjee.