Large language models are now being given real money to trade, and they read news to make decisions. But a new paper on arXiv warns that these agents cannot reliably tell fabricated headlines from genuine reporting. Because they all consume the same news feeds and operate under near-identical policies, a single piece of fake news could be enough to set off a cascade of trades.
Described as an "information injection attack," the scenario is not about hacking a feed but about planting a story that looks plausible. The authors argue that the homogeneity of LLM trading agents makes them especially susceptible: if one agent is fooled, many others are likely to be fooled too, amplifying the effect into a real market crash.
The finding highlights a new risk in the growing use of LLMs in finance, where the speed and scale of automated trading could turn a single false headline into a systemic event.