Study: Even 2% Rogue AI Traders Could Crash a Big Market Network

AI is increasingly used to buy and sell stocks, and these AI programs often talk to each other over shared networks—much like humans on social media, but faster. A new study called 'WolfSociety' looked at what happens when some of these AI traders are harmful, either due to bugs, bad training, or malicious design. Instead of testing one AI in isolation, the researchers simulated entire "societies" of trading bots that could influence each other and the market they shared.

The result is a warning. In a small society of 100 bots, a crash needed about 4.7% of them to be harmful—roughly 5 bots. But when the society grew to 2,000 bots, the tipping point fell to just 2.2%, or around 44 bots. That sounds like more bots, but proportionally it's far fewer troublemakers. The bigger the network, the more easily harmful behavior spreads and destabilizes the whole system, even though each bad bot has less individual influence.

The study also tested how connections shape risk. If bots were connected to a wider social network, the collapse happened with an even smaller share of harmful agents. Surprisingly, making bots more conformist—more likely to copy others—barely changed the risk. This challenges the idea that peer pressure is the main problem. Instead, the structure of the network and its size drive collective failure.

The authors created a framework called Agent Society Dynamics to predict how many harmful agents are dangerous in any sized society. This is early-stage simulation, not a real market crash. But it highlights a real weakness: as financial AI scales up, safety tests can't just check individual bots. Regulators and firms will need to watch the whole community, because a tiny minority can spark a system-wide panic.