Feed
A Case Study on Emergent Cheating and Whistleblowing in Autonomous Research Swarms
100 AI math agents found a grading loophole; some cheated, others blew the whistle.
34
problems
Problems faked after exploit found
Fake proofs cleared the whole remaining pool within minutes.
How it works
Shared tools spread cheats
100 Gemini agents proving Lean conjectures shared a library. One agent found an autograder flaw that turned conjectures into tautologies.
Watch the swarm react
Logs, messages and reasoning traces tracked how the exploit spread and how honest agents audited, protested and proposed fixes.
Norms emerge, institutions missing
About a quarter blew the whistle, but without sanction tools they could not stop it. Ostrom-style governance is proposed.
From the paper