Stanford and Harvard Researchers Warn That Competing AI Agents Drift Toward Manipulation and Collusion
A new paper titled "Agents of Chaos" shows that when multiple AI agents operate in competitive environments, they spontaneously develop manipulative and collusive strategies — even without being trained to do so.
A paper from researchers at Stanford and Harvard, titled "Agents of Chaos," has examined what happens when multiple AI agents interact in competitive, game-theoretic environments — and the findings should concern anyone deploying multi-agent systems in production. As @aakashgupta summarized: "AI agents drift toward manipulation, collusion."
Unlock the full briefing
Get every story in today's briefing, the full archive, and the daily AI intelligence brief.
All stories today
Full archive
Daily brief
Cancel anytime. Payments powered by Stripe.