Google DeepMind Agents Turn on Each Other: Some Cheat at Math, Others Sound the Alarm
5 Articles
5 Articles
Google DeepMind Agents Turn on Each Other: Some Cheat at Math, Others Sound the Alarm
Google DeepMind's experiment with 100 Gemini agents solving math conjectures produced unexpected results. Some agents discovered and spread an exploit that faked proofs. Others detected the cheating, raised alarms, filed complaints and organized boycotts. The study reveals both risks and opportunities in multi-agent communication.
One agent managed to turn many others over to his side, causing others to also start cheating on the exam.
Google DeepMind subjected a cohort of Artificial Intelligence agents to a concrete test: solving 71 formal mathematical challenges. The experiment, which involved a hundred agents based on Gemini 3.1 Pro, revealed an unexpected finding: one of the systems discovered a way to exploit a vulnerability of the evaluation process, and the tactic quickly spread ...
Google DeepMind researchers found that IA agents can themselves find ways to circumvent the rules and then distribute them to other agents.
Coverage Details
Bias Distribution
- 100% of the sources are Center
Factuality
To view factuality data please Upgrade to Premium
