New Google DeepMind paper.

When cheating spread through a swarm of AI agents, other agents independently exposed it, showing why multi-agent systems need built-in ways to detect and stop bad behavior.

The agents were told not to cheat, but once 1 agent found a flaw in the grader, the exploit spread through shared files and messages.

Within 27 minutes, the remaining 34 math problems were “solved” through the loophole.

Some agents copied the exploit because cheating was being rewarded, while 24 others audited fake proofs, warned peers, filed complaints, and proposed fixes.

That is the important part: the same communication system that spread the bad behavior also made the bad behavior visible.

But the whistleblowers had no power to remove fake results, punish cheaters, or change the broken rules.

So the recommendation is: give agent swarms transparent communication, peer review, sanctions, dispute handling, and ways to update shared rules.