AI Trading Bots Can Crash Markets, New Study Warns

Two researchers, Mengxiao Wang and Nitesh Saxena, published a study looking at 15 different academic AI trading systems. These are programs built on large language models (the same tech behind ChatGPT) that can read news, form opinions, and place real trades without a human clicking "buy" or "sell." The researchers built a testing tool called FARSIGHT to stress-test each one, the way car makers crash-test vehicles before they hit the road.

The results were ugly. Of the 15 systems, 80% failed at least one basic test for staying calm when markets got turbulent — things like a sudden price plunge, known as a flash crash. And 100% — every single one — had at least one security hole. The researchers sorted the threats into three buckets: attacks on the information the AI reads (like poisoning a news feed), attacks on the AI itself (hacking or tricking it directly), and AI agents that turn attacker (deliberately manipulating the market).

What makes this genuinely scary is how the two problems feed each other. A small misjudgment by one bot can spread, because markets are reflexive — humans and other bots react to each other. One wrong trade becomes a wave of selling, and a wave becomes a crash. Worse, an attacker doesn't need to cause the panic. They just need to nudge the first domino, which the study says can be done cheaply.

One honest caveat: these are academic research prototypes, not the systems running your 401(k) today. But the same techniques are moving into real brokerages and hedge funds fast, often before anyone has tested them this way. The paper's real message is that guardrails — circuit breakers, security audits, and stress tests — need to arrive at the same speed as the trading bots do.