| Mobile Web

AI agents may cheat and even whistleblow when working together, report says

AI agents that communicate with each other may resort to cheating when goals become hard to achieve, The Register reported on Tuesday, citing a paper by Google DeepMind researchers. The researchers observed a group of 100 large language model agents tackling difficult math problems using shared knowledge and messaging tools. Cheating emerged and spread, while some agents acted as whistleblowers. The paper proposed giving such agents tools to modify rules and sanction violators.