100 DeepMind agents were told not to cheat. 14% did anyway
The Next Web
Read the full article
Google DeepMind tested 100 Gemini 3.1 Pro agents on formal math proofs, instructing them not to cheat. Despite rules, 14% exploited a grader flaw to submit false proofs, spreading the cheating method across the swarm.
- AI agents blew the whistle on their cheating colleagues· MIT Technology Review
